Utah State University

DigitalCommons@USU
All Graduate Teses and Dissertations Graduate Studies, School of
7-1-2011
Document Flash Termography
Cory A. Larsen
Utah State University
Tis Tesis is brought to you for free and open access by the Graduate
Studies, School of at DigitalCommons@USU. It has been accepted for
inclusion in All Graduate Teses and Dissertations by an authorized
administrator of DigitalCommons@USU. For more information, please
contact digitalcommons@usu.edu.
Recommended Citation
Larsen, Cory A., "Document Flash Termography" (2011). All Graduate Teses and Dissertations. Paper 1018.
htp://digitalcommons.usu.edu/etd/1018
DOCUMENT FLASH THERMOGRAPHY
by
Cory A. Larsen
A thesis submitted in partial fulllment
of the requirements for the degree
of
MASTER OF SCIENCE
in
Electrical Engineering
Approved:
Dr. Doran J. Baker Dr. Gene A. Ware
Major Professor Committee Member
Dr. Todd Moon Dr. Mark R. McLellan
Committee Member Vice President for Research and
Dean of the School of Graduate Studies
UTAH STATE UNIVERSITY
Logan, Utah
2011
ii
Copyright c Cory A. Larsen 2011
All Rights Reserved
iii
Abstract
Document Flash Thermography
by
Cory A. Larsen, Master of Science
Utah State University, 2011
Major Professor: Dr. Doran J. Baker
Department: Electrical and Computer Engineering
This thesis presents the application of ash thermography techniques to the analysis of
documents. The motivation for this research is to develop the ability to non-destructively
reveal covered writings in archaeological artifacts such as the Codex Selden or Egyptian car-
tonnage. Current common signal processing techniques are evaluated for their eectiveness
in enhancing subsurface writings found within a set of test documents. These processing
techniques include: false colorization, contrast stretching, histogram equalization, median
ltering, Gaussian low-pass ltering, layered signal reconstruction and thermal signal recon-
struction (TSR), several contrast image denitions, dierential absolute contrast (DAC),
correlated contrast, derivative images, principal component thermography (PCT), dynamic
thermal tomography (DTT), pulse phase thermography (PPT), tting-correlation analy-
sis (FCA), Hough transform thermography (PTHTa), and transmission line matrix tting
algorithm (TLMFa). New processing techniques are developed and evaluated against the
existing techniques. The ability of ash thermography coupled with processing techniques
to reveal subsurface writings and document strikeouts is evaluated. Flash thermography
parameters are evaluated to determine most eective value for the document.
In summary, this thesis reports the following contributions to the existing scientic
knowledge:
iv
1. A comprehensive analysis of existing pulsed thermography processing techniques.
2. New pulsed thermography processing techniques that improve upon the results of the
existing techniques were developed.
3. A proof-of-concept for detecting subsurface ink writings in documents.
4. Veries the capability of pulsed thermography techniques to detect document strike-
outs.
5. Demonstrates the ability to enhance surface writings based on dierences in thermal
characteristics when optical characteristics do not dier signicantly.
6. Demonstrates that pulsed thermography signicantly improves upon multi-spectral
imaging for subsurface and surface writing enhancement.
7. Provides an evaluation of ash thermography parameters for the most eective doc-
ument imaging.
(146 pages)
v
Acknowledgments
Special thanks to Pedro Sevilla and James Peterson of the Space Dynamics Labora-
tory for the use of their equipment and helping with the acquisition of the thermal images.
Thank you to Dr. Gene A. Ware for his mentorship and guidance throughout the project.
Thanks to Dr. Doran J. Baker of the Rocky Mountain NASA Spacegrant Consortium for
funding the research and providing assistance as needed. Finally, thanks goes to Dr. Todd
Moon and Dr. Jake Gunther of Utah State University for sharing their advice and expertise.
Cory A. Larsen
vi
Contents
Page
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
Acronyms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 General Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Research Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4.1 Document Imaging and MSI . . . . . . . . . . . . . . . . . . . . . . 3
1.4.2 Flash Thermography Processing Algorithms . . . . . . . . . . . . . . 4
2 General Pulsed Thermography Considerations . . . . . . . . . . . . . . . . . . . . . 6
2.1 Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Instrumentation Requirements . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Flash Thermography Design Variables . . . . . . . . . . . . . . . . . . . . . 7
2.3.1 Thermal Imager Characteristics . . . . . . . . . . . . . . . . . . . . . 7
2.3.2 Thermal Properties of the Target . . . . . . . . . . . . . . . . . . . . 9
2.3.3 Characteristics of the Transmitting Medium . . . . . . . . . . . . . . 10
2.3.4 Excitation Pulse Characteristics . . . . . . . . . . . . . . . . . . . . 10
3 Experiment Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1 Instrumentation Used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Target Documents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.1 Subsurface Imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2.2 Strikeouts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2.3 Surface Ink Enhancement . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2.4 Egyptian Cartonnage . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4 Processing Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.1 Background Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.2 Existing Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.2.1 Pseudo-Color Images . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.2.2 Contrast Stretching . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.2.3 Histogram Equalization . . . . . . . . . . . . . . . . . . . . . . . . . 20
vii
4.2.4 Image Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.2.5 Synthetic Signal Reconstruction Techniques . . . . . . . . . . . . . . 21
4.2.6 Contrast Denitions . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.2.7 Dierential Absolute Contrast (DAC) . . . . . . . . . . . . . . . . . 24
4.2.8 Derivative Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.2.9 Principal Component Thermography (PCT) . . . . . . . . . . . . . . 26
4.2.10 Dynamic Thermal Tomography (DTT) . . . . . . . . . . . . . . . . . 27
4.2.11 Pulse Phase Thermography (PPT) . . . . . . . . . . . . . . . . . . . 28
4.2.12 Correlation Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.2.13 Transmission Line Matrix Fitting Algorithm (TLMFa) . . . . . . . . 31
4.2.14 Hough Transform Thermography (PTHTa) . . . . . . . . . . . . . . 33
4.3 Developed Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.3.1 Time-Dierence Contrast . . . . . . . . . . . . . . . . . . . . . . . . 34
4.3.2 Total Harmonic Distortion (THD) . . . . . . . . . . . . . . . . . . . 35
4.3.3 Markov Error Contrast (MEC) . . . . . . . . . . . . . . . . . . . . . 36
4.3.4 Time Constant Analysis (TCA) . . . . . . . . . . . . . . . . . . . . . 37
4.3.5 Signal Detection and Matched Filtering (MF) . . . . . . . . . . . . . 37
4.3.6 Convex Optimization Signal Detection Technique . . . . . . . . . . . 41
5 Analysis of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.1 Processing Techniques Results . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.1.1 Subsurface Inks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.1.2 Document Strikeouts . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.2 Comparison of Algorithm Wall Times . . . . . . . . . . . . . . . . . . . . . 62
5.3 Comparison with Multi-Spectral Imaging (MSI) . . . . . . . . . . . . . . . . 62
5.3.1 Subsurface Inks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.3.2 Document Strikeouts . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3.3 Surface Writing Enhancement . . . . . . . . . . . . . . . . . . . . . . 67
5.4 Feasibility of Application to Archaeological Artifacts . . . . . . . . . . . . . 69
5.4.1 Egyptian Cartonnage . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.4.2 Codex Selden . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6 Conclusions and Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.1 Thermography Data Acquisition Parameters . . . . . . . . . . . . . . . . . . 77
6.2 Pre-Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.3 Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.4 Post-Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.5 Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.6 Signicant Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.7 Recommendations for Future Research . . . . . . . . . . . . . . . . . . . . . 81
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Appendix A Pulsed Thermography Toolbox . . . . . . . . . . . . . . . . . . . 93
A.1 Pulsed Thermography Toolbox (PTT) Introduction . . . . . . . . . 93
A.2 Load Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
viii
A.3 Manually Manipulate Data Panel . . . . . . . . . . . . . . . . . . . . 95
A.4 Pre-Processing Panel . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
A.5 Processing Panel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
A.6 Post-Processing Panel . . . . . . . . . . . . . . . . . . . . . . . . . . 98
A.7 Other GUI Sub-Components . . . . . . . . . . . . . . . . . . . . . . 98
A.8 Explore Data Sub-GUI . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Appendix B Code Listings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
B.1 Contrast Denitions & Dierential Absolute Contrast (DAC) . . . . 101
B.2 Thermal Signal Reconstruction (TSR) and Derivative Images . . . . 106
B.3 Principal Component Thermography (PCT) . . . . . . . . . . . . . . 111
B.4 Dynamic Thermal Tomography (DTT) . . . . . . . . . . . . . . . . . 113
B.5 Pulse Phase Thermography (PPT) . . . . . . . . . . . . . . . . . . . 118
B.6 Time-Dierence Contrast (TDC) . . . . . . . . . . . . . . . . . . . . 120
B.7 Total Harmonic Distortion (THD) . . . . . . . . . . . . . . . . . . . 122
B.8 Markov Error Contrast (MEC) . . . . . . . . . . . . . . . . . . . . . 125
B.9 Time Constant Analysis (TCA) . . . . . . . . . . . . . . . . . . . . . 127
B.10 Matched Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
ix
List of Tables
Table Page
5.1 Signal reconstruction wall times. . . . . . . . . . . . . . . . . . . . . . . . . 46
5.2 Algorithm wall times. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.1 Algorithm rating system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.2 Processing algorithms summarized. . . . . . . . . . . . . . . . . . . . . . . . 82
x
List of Figures
Figure Page
2.1 Flash thermography setup. . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Document structure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1 Pulsed thermography test system. . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Egyptian cartonnage. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
5.1 Test document 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.2 Noise reduced images at t = 228.8 ms. . . . . . . . . . . . . . . . . . . . . . 45
5.3 Pseudo-color image at t = 228.8 ms. . . . . . . . . . . . . . . . . . . . . . . 46
5.4 Contrast stretching at t = 228.8 ms. . . . . . . . . . . . . . . . . . . . . . . 47
5.5 Dierential absolute contrast (DAC) images. . . . . . . . . . . . . . . . . . 47
5.6 Contrast denitions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.7 Derivative images. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.8 Principal component thermography (PCT) images. . . . . . . . . . . . . . . 49
5.9 Dynamic thermal tomography (DTT) images. . . . . . . . . . . . . . . . . . 50
5.10 Pulse phase thermography (PPT) image at 1.2 Hz. . . . . . . . . . . . . . . 50
5.11 Correlation images. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.12 Transmission line matrix tting algorithm (TLMFa) image. . . . . . . . . . 51
5.13 Hough transform coecient images. . . . . . . . . . . . . . . . . . . . . . . 52
5.14 Time-dierence contrast images. . . . . . . . . . . . . . . . . . . . . . . . . 52
5.15 Total harmonic distortion (THD) images. . . . . . . . . . . . . . . . . . . . 53
5.16 Markov error contrast (MEC) images. . . . . . . . . . . . . . . . . . . . . . 54
5.17 Time constant analysis (TCA) image. . . . . . . . . . . . . . . . . . . . . . 54
xi
5.18 Matched lter images. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.19 Strikeouts test set. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.20 Contrast denition strikeout images. . . . . . . . . . . . . . . . . . . . . . . 57
5.21 DAC strikeout image with t

= 11.5 ms and t = 57.5 ms. . . . . . . . . . . . 57

5.22 DTT classical maxigram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.23 Markov error contrast (MEC) strikeout image. . . . . . . . . . . . . . . . . 59
5.24 Principal component thermography (PCT) strikeout images. . . . . . . . . . 59
5.25 Pulse phase thermography (PPT) amplitude strikeout image. . . . . . . . . 59
5.26 Transmission line matrix tting algorithm (TLMFa) strikeout image. . . . . 60
5.27 Time constant analysis (TCA) strikeout image. . . . . . . . . . . . . . . . . 61
5.28 Time-dierence contrast strikeout images. . . . . . . . . . . . . . . . . . . . 61
5.29 Strikeout derivative images . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.30 Algorithm wall times. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.31 MSI vs. FT strikeout results. . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.32 Cardstock test document under-layers. . . . . . . . . . . . . . . . . . . . . . 65
5.33 MSI vs. PT cardstock results. . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.34 MSI vs. FT cardstock 2 results. . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.35 Papyrus test document. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.36 MSI vs. FT papyrus results. . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.37 Paint test document. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.38 MSI vs. FT painted images. . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.39 MSI vs. FT gesso images. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.40 MSI vs. FT strikeout results. . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.41 MSI vs. FT red surface results. . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.42 MSI vs. FT cartonnage images. . . . . . . . . . . . . . . . . . . . . . . . . . 72
xii
5.43 Pre-ash FT cartonnage image. . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.44 MSI vs. FT Codex Selden images data set 1. . . . . . . . . . . . . . . . . . 74
5.45 MSI vs. FT Codex Selden images data set 2. . . . . . . . . . . . . . . . . . 75
5.46 MSI vs. FT Codex Selden images data set 3. . . . . . . . . . . . . . . . . . 76
A.1 Pulsed Thermography Toolbox main GUI. . . . . . . . . . . . . . . . . . . . 94
A.2 Load images panel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
A.3 Select les interface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
A.4 Display panel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
A.5 Manually Manipulate Data Panel. . . . . . . . . . . . . . . . . . . . . . . . 96
A.6 Pre-Processing Panel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
A.7 Processing Panel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
A.8 Post-Processing Panel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
A.9 Explore Data Sub-GUI. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
xiii
Acronyms
NDT&E Non-destructive Testing & Evaluation
MSI Multi-Spectral Imaging
IR Infrared
PDF Probability Distribution Function
DAC Dierential Absolute Contrast
TSR Thermal Signal Reconstruction
PCT Principal Component Thermography
PCM Primary Contrast Mode
DTT Dynamic Thermal Tomography
PPT Pulse Phase Thermography
FCA Fitting Correlation Analysis
PTHTa Pulsed Thermography Hough Transform Algorithm
TLM Transmission Line Matrix
TLMFa Transmission Line Matrix Fitting Algorithm
TDC Time-Dierence Contrast
MF Matched Filtering
THD Total Harmonic Distortion
MEC Markov Error Contrast
TCA Time Constant Analysis
GUI Graphical User Interface
NETD Noise Equivalent Temperature Dierence
SNR Signal-to-Noise Ratio
MF Matched Filter
SAM Spectral Angle Map
ACE Adaptive Coherence Estimator
1
Chapter 1
Introduction
1.1 General Background
Flash thermography, a subset of pulsed thermography or pulsed video thermography,
is a technique used for non-destructive testing and evaluation (NDT&E) in a variety of
materials, including concrete [13], high-density polyethylene [4], aerospace composites [5,6],
wood and wood-based materials [7], and adhesive bond evaluation [8,9]. Flash thermography
uses optical ash lamps to inject heat energy into a material. A high-speed infrared camera
records the temperature of the material surface as the heat energy diuses through the
material. The video sequence is then processed to enhance the contrast of relatively warm
(or cold) areas on the surface that result from thermal reections caused by material aws.
Flash thermography was yet to be applied to documents or archaeological artifacts to
reveal covered writings; for instance those found in the Codex Selden or Egyptian cartonnage
[10]. The research presented in this thesis was used to develop the theory and application
of ash thermography to documents, and lays a foundation for the application of this
technology to archaeological artifacts in general. Development of ash thermography for this
application provides the capability to non-destructively reveal covered writings advancing
the knowledge about ancient cultures without damaging irreplaceable artifacts.
1.2 Problem Statement
Motivation for this research derives from the desire to analyze ancient archaeological
documents with a non-destructive approach. Specically of interest is the imaging of subsur-
face writings that may be obscured with a layer of some type of material. Example ancient
documents this technology could be applied to include, but are not limited to: palimpsests
2
from the Roman Empire, Mesoamerican codices, and Egyptian cartonnage. Another inter-
est is the use of ash thermography for detection of textual changes, such as strikeouts,
where older ink writings are covered with a more recent layer of ink writing. Finally, the
ability of ash thermography to enhance contrast of surface writings through dierences in
thermal characteristics, rather than optical characteristics, are of interest.
1.3 Research Objectives
This thesis investigates the application of ash thermography to the textual analysis of
documents. Areas of interest include revealing subsurface layers of writing, detecting textual
changes such as document strikeouts, and enhancing surface writing through dierences
in materials properties between the ink and the document. A strong engineering basis
is provided to facilitate the implementation of the investigations herein reported on other
documents and archaeological artifacts. The scope of this research is limited to single-sided,
optical ash thermography techniques and the corresponding processing algorithms.
Representative processing algorithms were chosen from the large body of ash thermog-
raphy literature and applied to ash thermography images of documents with the eective-
ness of each algorithm ascertained. In addition, unique new algorithms were investigated.
This resulted in a determination of document types for eective application of ash ther-
mography and a toolbox of processing techniques was collated.
A study was performed to evaluate the eectiveness of ash thermography in compari-
son with the multi-spectral imaging (MSI) standard. Attempts to validate the results were
made using samples of Egyptian cartonnage and the Codex Selden. The results obtained
are an important contribution to assist anthropology eorts in the non-destructive study
of ancient documents. The accompanying graphics could provide signicant value in the
analysis and understanding of ancient cultures and societies.
1.4 Literature Review
The following literature review is split into two subsections which focus upon: (1)
document imaging techniques, and (2) ash thermography processing techniques.
3
1.4.1 Document Imaging and MSI
Several imaging techniques have been developed to enhance the observation of surface
writing and under-writing contained in archaeological artifacts. Infrared (IR) reectography
has been used in the analysis of paintings [11, 12] and papyrus [13]. The main application
of IR reectography is to view under-drawings beneath a layer of paint. With IR reectog-
raphy, a constant light source is used to excite the material. An IR imager captures an IR
reectogram detailing the dierent optical properties of the overlaying paint and the under-
writing [14]. In addition to IR reectography, transient thermographic techniques, including
pulsed thermography, have also been used to analyze paint layers in artwork such as fres-
coes [15, 16] and general artwork [1720]. A comparative study was performed comparing
pulse thermography, lateral heating thermography, and modulated thermography for the
analysis of frescoes [16]. Pulsed thermography was shown to be successful in areas where
X-radiography, infrared reectography, and UV examination had been unsuccessful [17].
Pulsed thermography has not yet been applied to the evaluation of ancient documents or
other types of archaeological artifacts.
A common technique for analyzing ancient documents is multi-spectral imaging (MSI).
MSI has been shown to be eective for enhancing contrast between underwriting, overwrit-
ing, and the document substrate [21,22]. MSI is performed by imaging documents in narrow
spectral bands of light, allowing the spectral signature of the dierent document materials
to be evaluated. Processing techniques for MSI images include, but are not limited to,
the use of Markov random elds [21], spectral clustering [23], principal and independent
component analysis [22], and linear spectral mixture analysis [22, 24].
The use of MSI has been successful in revealing obscured writing on the Archimedes
palimpsest [24], carbonized scrolls [23], oxyrhynchus papyri [25], and the dead sea scrolls [26,
27]. MSI is most eective in enhancing the observation of writing that appears on or near
the surface of the document. For example, the eectiveness of MSI to reveal the under
codex within the Codex Selden was shown to be limited [28]. Another technique currently
under investigation is X-Ray Fluorescence Imaging (XRF) [29].
4
1.4.2 Flash Thermography Processing Algorithms
A large body of processing techniques exist for enhancing defect contrast in pulsed
thermography data. A thorough investigation of the literature revealed several techniques
used for processing ash thermography data. Techniques include enhanced visualization
through pseudo-color images [30, 31] and contrast stretching [32]. Spatial image noise
reduction include techniques such as median ltering [30, 31, 33] and Gaussian low-pass
ltering [30, 31, 33]. Synthetic signal reconstruction techniques are applied to reduce tem-
poral noise and include layered signal reconstruction [34] and thermal signal reconstruc-
tion (TSR) [35, 36]. Several contrast image techniques include four contrast image deni-
tions [30], dierential absolute contrast (DAC) [3739], and interpolated dierential absolute
contrast (IDAC) [40], as well as several similar techniques [4143]. Analogies to modulated
thermography techniques are made through pulse phase thermography (PPT) [4450].
Other common processing routines include derivative images [33,35,36,45,5155], prin-
cipal component thermography (PCT) [56, 57], and dynamic thermal tomography (DTT)
[5860]. Correlation images [61, 62] are obtained by measuring the correlation coecient
between a measured and a desired signal. Hough transform thermography (PTHTa) [63, 64]
uses a parametric transform to aid in aw detection. The transmission line matrix tting
algorithm (TLMFa) [65, 66] models thermal diusion through transmission line theory to
characterize material aws.
Techniques that account for lateral diusion include inverse scattering techniques [67
69], pulse-echo thermal imaging [70, 71], and point spread functions [72]. Other techniques
include neural networks [30, 58, 73], Laplace transform techniques [64, 74] based on thermal
quadropole theory [43, 75], aw detection thermal tomography [58, 76], nonlinear tting and
optimization methods [58, 77, 78], adaptive thermal tomography [58, 79], and learning ma-
chines [80]. Image aw segmentation [81] routines have been explored. In addition, several
5
tomographic techniques were developed based o the Algebraic Reconstruction Technique
(ART) [8284]; however, a raster-scan of the target using a laser for excitation is required.
6
Chapter 2
General Pulsed Thermography Considerations
2.1 Terminology
Throughout this thesis, the term defect is used in a manner consistent with current
NDT&E literature. For applying ash thermography to documents, it is assumed that the
ink under-writing can be treated as a defect in the document. Therefore, the terms defect
and ink are used interchangeably throughout this research.
2.2 Instrumentation Requirements
Basic instrumentation requirements for performing pulsed thermography include a ther-
mal viewer, an excitation source, and data collection and processing hardware [85]. Typical
thermal imagers used in ash thermography applications are high-speed, mid-IR cameras.
Using IR cameras, only qualitative thermograms are possible. Quantitative thermograms
can be obtained using a radiometer; however, the present research is concerned only with the
relative temperature dierences between the area of interest and the rest of the document.
An excitation source is required for energy injection into the target sample. It is desired
to generate a uniform thermal excitation across the entire surface of the sample. Pulsed
sources include ash lamps and pulsed lasers. The present scope of work is limited to optical
ash lamps as excitation sources.
The data collection and processing hardware usually consists of a computer to store
the data and perform additional processing. However, some cameras include real-time
processing within the camera itself. There are several existing software packages available
for processing thermal data such as the Altair-Li suite provided by Cedip, RTools by FLIR,
the ThermoFitPro software by Innovation Inc., and the open-source IR View Toolbox and
GUI for Matlab [86].
7
A typical ash thermography setup is presented in Figure 2.1. Optical ash lamps
provide the necessary energy which the document absorbs as heat and conducts through
the document. Infrared energy is radiated from the surface and recorded by a high-speed
infrared camera. In document sections containing ink underwriting, heat is reected back
to the surface creating a warm spot. The ash thermography video sequence is sent to a
computer to process the images to increase defect visualization.
2.3 Flash Thermography Design Variables
Numerous considerations must be taken into account when applying ash thermogra-
phy. These variables fall within four broad categories: 1) Characteristics of the thermal
imager, 2) Thermal properties of target, 3) Characteristics of the transmitting medium, 4)
Excitation source characteristics. Each of these categories are summarized [85].
2.3.1 Thermal Imager Characteristics
Evaluation characteristics of the thermal imager include:
Temperature Sensitivity: Also referred to as the minimum resolvable temperature or
noise-equivalent temperature dierence (NETD) [87], temperature sensitivity is the
measure of the smallest temperature dierences a camera is able to detect. The tem-
perature sensitivity of the camera must be greater than the dierential temperature
Fig. 2.1: Flash thermography setup.
8
signal indicating the dierence in temperature between a defect-free region and a
defect [88].
Image Acquisition Rate: The required image acquisition rate, also referred to as sam-
pling rate or video frame rate, is dependent on the thermal diusivity of the target
material. A higher material thermal diusivity requires a faster frame rate to prop-
erly capture the transient response of the thermal diusion through the document.
Dening as the time constant of the system, it can be assumed that a sampling rate
of f
10

is sucient [88].
Image Spatial Resolution: The image spatial resolution is a measure of the physical
area an individual pixel views. A defect must have an area of at least one pixel in
order to be detectable. Generally, the more pixels containing the defect the easier
detection becomes.
Spectral Range: Spectral range species the wavelengths the camera is able to image.
For ash thermography, the desired spectral range is either 2-5 m or 8-12 m. The
spectral range of a camera can be limited using a bandpass lter [85].
Temperature Range: The temperature ranges determine the maximum and minimum
temperatures that are detectable. This is important when there are large temperature
dierences due to defects or other objects with large dierences in material properties
in the eld of view.
Total Field of View: The total eld of view determines the area the camera is able to
image at a time. The eld of view must be large enough to image the region of interest.
Sensor Environment: The sensor environment includes the environmental properties of
the experiment setup such as atmospheric temperature, pressure, humidity, or any
other environmental variables that may aect the cameras ability to image properly.
9
2.3.2 Thermal Properties of the Target
Certain thermal properties of the target surface will aect the eectiveness of ash
thermography. Properties to consider regarding the target surface include:
Thermal Emissivity: The thermal emissivity is a measure of the ability of the surface
material to exchange thermal energy with its surroundings. Targets with high and
uniform surface emissivity are most eective. The eects of surface emissivity non-
uniformities can be reduced by dierencing a pre-ash image from the data set. In
addition, the surface can be coated with washable black paint to increase emissivity,
although this is unfeasible for documents or archaeological artifacts [89].
Thermal Reectivity: The reectivity of the surface determines the amount of the ini-
tial energy impulse that will be absorbed. Low thermal reecitivities indicate high
amounts of energy absorption into the document from the excitation pulse. A target
with a high reectivity may introduce artifacts in the acquired images created by re-
ections from gray bodies surrounding the target. The thermal emissivity, reectivity,
and transmitivity of a surface are related through
+ + = 1, (2.1)
where is the thermal eusivity, is the reectivity, and is the transmitivity [85].
Thermal Diusivity: Thermal diusivity is inversely proportional to the time constant
for thermal diusion through the target. The higher the thermal diusivity, the less
time it takes for the energy to diuse through the document, thus requiring faster
sampling rates (see Image Acquisition Rate). The lower the thermal diusivity, the
longer the total sampling time required. The total sampling time required to allow
for diusion to the back wall is specied by [90, 91]
t
exp
=
L
2

, (2.2)
10
where L is the thickness of the layer and is the thermal diusivity of the material.
2.3.3 Characteristics of the Transmitting Medium
Characteristics of the medium between the source excitation pulse and the target must
be considered. In most cases the transmitting medium is air. In air, for short distances up
to a few meters, these characteristics can be largely ignored. In addition, when imaging in
the 3-5 m range air is relatively free of signicant spectral losses [85].
2.3.4 Excitation Pulse Characteristics
In this text, only pulsed excitation sources are considered. Characteristics of the input
pulse such as amplitude, shape, and timing are now discussed.
Pulse Amplitude: The maximum amount of energy able to be inputted into the target
has an upper limit determined by the temperature at which the material will begin to
be damaged. This is checked when the sample temperature is maximal thus limiting
either the amplitude or the duration of the excitation [88].
In order to detect a defect, or the back wall of the document structure, (Figure 2.2), the
signal must be above the Noise Equivalent Temperature Dierence (NETD), or stated
another way, the temperature signal of a defect must have a signal-to-noise (SNR) ratio
greater than 1 [88]. It is customary to dene the minimum signal detection level to be
a multiple of the NETD, with a rule of thumb SNR value of n = 2. The temperature
dierence needed to detect the back wall of the target is given as [87]
T
wall
= n, (2.3)
where = NETD and n is the SNR value. The minimum energy required for detecting
the back wall is given by
Q
min
= nCL, (2.4)
11
Fig. 2.2: Document structure.
where is the density, C is the specic heat, and L is the thickness of the layer. Note
that the thermal diusivity does not aect the total energy requirement, however it
does aect the running time of the experiment (see Thermal Diusivity).
In order to observe a defect, the signal detection level must be an additional
factor greater than the back wall detection level. Thus for a given signal level, m,
and a detectability threshold, n, the maximum defect depth able to be detected is
given by [87]
d
max
=
m
m+n
L, m n. (2.5)
At the minimum energy level, Q
min
, the deepest detectable defect is at a depth of
L
2
.
The ability to detect deeper defects increases at a logarithmic rate with energy levels.
The amount of thermal energy absorbed by the target is dependent on several
parameters, with a best case scenario found to be approximately 25% eciency. There-
fore, the ash electrical energy must be [87]
W
electrical

Q
min
Area
Eciency
. (2.6)
Note that these detection levels are for an idealized situation and represent a lower
limit on energy requirements. However, the processing techniques discussed in Chap-
ter 5 may facilitate defect detection up to and beyond these limitations.
12
Pulse Shape: The energy pulses inputted into the target are modeled as an impulse func-
tion or as a rectangular pulse. However, since the optical pulses are generated from
a capacitor discharge, an exponentially decaying tail continues to input energy. This
energy tail can obscure the rst few frames of the sample surface. It was found that by
using a ash controller to shorten the length of the tail, better temperature responses
were obtained [92, 93]. However, the shortening the tail reduces the amount of energy
input into the target creating a trade o. The best results are obtained using the
shortest pulse possible but still providing sucient energy to the system. For low
thermally conducive materials, it was shown that using a double pulse [94] or a train
of pulses [95] could improve the signal-to-noise ratio leading to improved imaging.
Pulse Timing: Correcting for the timing oset between acquisition time and frame time
increases correlation between modeled and experimental results [92]. The camera
integration period can have a dramatic eect on the rst frame. It was found that
this eect could be minimized by using an integration period of less than the frame
read-out period in combination with using the sample time measured from the heat
pulse to the center of the integration period [92]. These pulse timing methods can
often be dicult to accomplish in practice.
13
Chapter 3
Experiment Setup
3.1 Instrumentation Used
For the experiments performed in this thesis, a Lockheed-Martin/Santa Barbara Fo-
cal plane model SBF 180 thermal camera system was used with a custom data collection
computer and software. The excitation energy pulses were provided by four SunPak Pro-
System 622 Super optical ash units stationed on the sides of the target document. The
camera was set horizontally on a table and aimed at an angled mirror. This mirror was
used to image the test document laid at on the surface of a table. Two camera lters were
tested, a 3.42 4.05 m and a 2.65 3.24 m bandpass lter. Figure 3.1 shows the pulsed
thermography test system.
3.2 Target Documents
Test documents were created to simulate challenges that occur in imaging ancient doc-
uments and other documents of interest. These documents fall into three broad categories:
(1) subsurface imaging, (2) strikeouts, and (3) surface ink enhancement. Subsurface imaging
Fig. 3.1: Pulsed thermography test system.
14
refers to an attempt to image beneath a covering layer. Strikeouts refer to textual changes
made to the surface of the target document where two layers of ink overlap. The last cat-
egory, surface ink enhancement, refers to the attempt to enhance contrast of surface inks
through dierences in thermal characteristics between the ink and the document backing.
Finally, a segment of Egyptian cartonnage was used to validate the method on an artifact
of interest. Each of the document types are now discussed in detail.
3.2.1 Subsurface Imaging
There are several covering layers that are of interest to image beneath including pa-
pers, papyrus, a layer of mineral gesso, or a layer of paint. Imaging through the covering
layer allows the ability to non-destructively reveal subsurface writing that cannot be ob-
served without permanently damaging the document. Common documents of interest in
this category are Egyptian cartonnage or pages in ancient manuscripts that have become
permanently stuck together.
The mineral gesso covering layer contains a mixture of acrylics and calcium carbon-
ate powder. This is motivated by documents such as the Codex Selden, a Mesoamerican
palimpsest where the original codex was covered by a mineral gesso layer and a new codex
drawn on the surface.
Imaging through a layer of paint will facilitate imaging Egyptian cartonnage, where old
papyrus was used in the mummication process and decoratively painted over the existing
writing. Other applications include imaging art works and frescoes.
Test documents were created using three inks applied in a pattern. The inks used
were carbon-based, iron gall, and ball-point pen. Covering layers consisted of papyrus, card
stock, gesso mixture, and oil-based paints.
3.2.2 Strikeouts
A strikeout test document was used to evaluate whether ash thermography techniques
can detect textual changes within a set of writing. It is common in (ancient) documents for
changes to be made to the original writing. MSI can be used to detect two layers of writing
15
that contains two dierent inks, but is not as eective when the over- and under-inks are
spectrally similar. Dual layers of a carbon-based ink, an iron gall ink, a ball-point pen ink,
and pencil lead were tested. In addition, a layer of white-out was used to determine the
ability to image changes in modern documents and a layer of indentations were used to test
surface inhomogeneities.
3.2.3 Surface Ink Enhancement
Surface ink enhancement is an attempt to reveal surface inks that are spectrally similar
to the document backing. Detection using ash thermography may enhance the contrast of
the inks with the surface through dierences in material thermal properties. Ancient iron
gall inks are known to fade in the infrared and lack signicant spectral dierences for MSI to
properly enhance. The test document created consisted of a thin red and blue ink on a red
card stock backing. Since the red and blue inks should be thermally similar, but spectrally
dierent, it allows for a simple comparison between enhancement using ash thermography
and MSI.
3.2.4 Egyptian Cartonnage
Flash thermography will be applied to a piece of Egyptian cartonnage. Used by the
ancient Egyptians in the mummication process, cartonnage consists of scraps of papyrus or
linen combined with a lime plaster similar to how modern papier-mache is used. It was then
molded into a desired shape and allowed to dry before being decoratively painted. Many
of the scraps of papyrus used contained ancient writing that is of interest to scholars. The
ash thermography techniques developed through using the test documents was applied
to the Egyptian cartonnage in an attempt to reveal such hidden writing. The cartonnage
provides a challenge for imaging because of its complex structure. Writings of interest may
be buried beneath a layer of paint, gesso, papyrus, or a combination of each. The piece of
cartonnage imaged is shown in Figure 3.2.
16
Fig. 3.2: Egyptian cartonnage.
17
Chapter 4
Processing Techniques
4.1 Background Theory
The majority of the algorithms herein presented process the image data on a pixel
by pixel basis, evaluating the time series of each pixel separately without accounting for
lateral diusion. These time series represent the post-ash surface temperature decay of
the material through time. It is often assumed that the diusion into the material is
signicantly greater than the lateral diusion and therefore the lateral diusion can be
neglected. This allows the diusion into the document to be described using the equation
for one-dimensional thermal diusion as given by

2
T
x
2
=
1

T
t
, (4.1)
where T is the temperature and is the thermal diusivity of the material. For an ideal
impulsive heat ux, the response for a semi-innite surface is [96]
T(x, t) =
Q
e

t
e

x
2
4t
, (4.2)
where e =

kc is the thermal eusivity of the material as determined by the thermal

conductivity, k, mass density , and specic heat c. Q is the quantity of energy absorbed
by the surface. Time is represented by t and the depth into the material is given by x.
Since the thermal imager can only respond to surface temperatures, (4.2) is evaluated at
x = 0, resulting in the surface temperature decay modeled by
T
surf
(t) = T(0, t) =
Q
e

t
. (4.3)
18
However, the thermal imager only gives relative temperatures, therefore let T
surf
(t) =
T
surf
(t) T
ambient
, where T
ambient
is the pre-ash initial temperature of the sample. The
response is thus more accurately described relative to the thermal imager as
T
surf
(t) =
Q
e

t
. (4.4)
Equation (4.4) provides a basis for many of the algorithms discussed in this section. This
equation can be further simplied to [54]
T
surf
(t) = T
init
_
T
s
t
, (4.5)
where T
init
is the value of the surface temperature at one time step, T
s
, and is given by
T
init
=
Q
e

T
s
. (4.6)
The techniques described throughout this section were implemented in MATLAB 2010b
and a graphical user interface (GUI) was developed to aid in the implementation of the
processing techniques. A tutorial for the GUI is given in Appendix A and the main algorithm
code is given in Appendix B.
4.2 Existing Techniques
The theories behind several existing techniques used to enhance visibility and charac-
terize material defects are outlined in the following sections.
4.2.1 Pseudo-Color Images
Human eyes contain two classes of light receptors: cones and rods. Cones are highly
sensitive to color and are able to discern smaller changes in intensity than rods. This is
due to the fact that cones, unlike rods, are connected to an individual nerve ending [31].
Therefore, it is useful to provide pseudo-color to the gray-scale intensity images obtained
from the thermal cameras. These pseudo-color images are obtained through a process of
19
intensity slicing and color coding. Intensity slicing is a technique used to separate dierent
intensity values in a gray-scale image. Given a gray-scale image with L intensity levels, the
image can be separated into P intensity levels where 0 < P < L 1. By color coding the
intensity values between regions, a pseudo-color image is created [31]. To avoid creating
a mosaic eect within the image, it is better to use a continuous color map. However in
some situations, choosing a discrete color mapping scheme may enhance the visibility of a
particular feature [30]. Pseudo-color images can be created in MATLAB using the colormap
command using the jet color map parameter.
Pseudo-colorization can be used as an aid to help enhance visualization of defects.
Pseudo-colorization is often used as a pre-processing technique to analyze the data before
processing. If no defects can be detected, even subtly, from a visual inspection then it is
usually an indicator that the processing algorithms will not be able to eectively enhance
any under writing. Pseudo-color images can also be used to aid in visualization of the
post-processed images.
4.2.2 Contrast Stretching
Contrast stretching is a point processing technique used to expand the dynamic range
of an image to increase visibility of image features. This technique is used to enhance
visibility of the raw thermal images or as a post-processing step for the other techniques
described in this section. The simplest form of contrast stretching is image normalization
which is given by
I
N
= (I c)
_
b a
d c
_
+a, (4.7)
where I is the input image with initial range [c, d] and I
N
is the normalized output image
in the desired range [a, b]. Note that the normalization process is used on individual frames
within the time sequence. Image normalization is greatly aected by dead pixels and other
outliers in pixel values. To compensate, a useful technique is to saturate the top 1% and
bottom 1% of pixels. This can be implemented with the MATLAB function imadjust.
20
4.2.3 Histogram Equalization
A histogram of a digital image is a discrete function that counts the number of pixels
within given intensity level bins. The histogram is a discrete estimation of the probability
density function (PDF) of the image. In histogram equalization, the goal is to transform the
histogram of the input image into an output image with a uniformly distributed histogram.
This technique enhances the global contrast of an image by adjusting the distribution of
intensities within an image. If a defect lies in a region with low local contrast, histogram
equalization will increase the contrast of that region allowing for enhanced defect visualiza-
tion.
The transformation on each pixel intensity value is represented by
s = T(r), (4.8)
where s is the equalized pixel value, r is the input pixel value, and T(r) represents the
transformation performed on r to obtain s. The transformation is given by [31]
s = T(r) = (L 1)
_
r
0
p
r
(w)dw, (4.9)
where L represents the number of intensity levels and w is a dummy variable of integration.
Equation (4.9) represents the cumulative distribution function of the image. It is proved
elsewhere that this transform results in a uniform distribution [31].
For discrete digital images, the transformation function becomes
s
k
= T(r
k
) =
L 1
MN
k

j=0
n
j
k = 0, 1, 2, ..., L 1 (4.10)
where MN is the total number of pixels in the image, and n
k
is the number of pixels that
have intensity r
k
. A plot of n
k
versus r
k
results in the histogram of the image [31].
Since the histogram is a discrete approximation of the PDF and no new intensity
levels can be created, perfectly at histograms are rare in practical images [31]. Histogram
21
equalization can be implemented in MATLAB using the histeq function. The histogram
equalization process is used to post-process the images after running the other algorithms
described in this section or to better evaluate the raw, unprocessed images.
4.2.4 Image Filters
Two image lters have been found to be useful in noise reduction in ash thermography
data, namely a median lter [31] and a Gaussian low-pass lter [30]. Median lters are useful
for removing salt and pepper noise. The median lter is a nonlinear lter which ranks the
pixels within a neighborhood, replacing the center pixel with the median of the intensity
values. The median lter has advantages over a mean lter because it is not aected by
outliers (such as those caused by dead pixels) and better preserves edges within the image.
The median lter is common in imaging software and can be implemented in MATLAB
using the medfilt2 function.
The Gaussian lter is a frequency-domain lter that assumes the image has a limited
bandwidth and any spatial frequencies above the given bandwidth are the result of noise.
Since image noise tends to be characterized by high spatial frequencies, a low-pass lter
can be used to reduce the noise content in the image. A derivation of the Gaussian lter
for thermal images can be found elsewhere [30]. The Gaussian lter can be implemented
in MATLAB using the fspecial command to create the lter and the imfilter command
to apply the lter to the image.
4.2.5 Synthetic Signal Reconstruction Techniques
Two synthetic signal techniques were employed to reconstruct the signal, namely ther-
mal signal reconstruction (TSR) [36, 52] and layered reconstruction (LR) [34]. Advantages
of using reconstructed, synthetic data include (1) signicant improvements in sensitivity,
(2) a reduction of blurring, (3) increased depth range, (4) decreased memory requirements,
and (5) improvements in signal-to-noise performance.
The TSR reconstruction processes each individual pixels time sequence, rather than
each frame as a whole. The time response data can be linearized by transforming to a
22
logarithmic domain. The logarithmic transform of (4.4) is
ln (T
surf
(t)) = ln (
Q
e
)
1
2
ln (t). (4.11)
This implies that, regardless of the thermal properties of the material, the logarithmic
decay response will be a straight line with a slope of
1
2
for an ideal, defect-free region.
The linearized sequence can be modeled using a least-squares t to a Nth order polynomial
ln[T
surf
(t)] =
N

n=0
a
n
[ln(t)]
n
. (4.12)
It was found that a fth or sixth order polynomial eectively acts as a low-pass lter to
smooth the data without reconstructing the noise [52]. The data can be reconstructed
using [36, 52]
T
surf
= exp
_
N

n=0
a
n
[ln(t)]
n
_
. (4.13)
In the layered reconstruction (LR) [34] algorithm, a multi-layered approach is taken.
It is proposed that the signal can be reconstructed using the following equation:
T
surf
(t) = T
f
+
j

i=1
A
i
e

i
, (4.14)
where A
i
is the amplitude of each exponential function and T
f
is the steady-state temper-
ature. The time constant of each layer is dened as
i
=
1

i
B
2
i
, where
i
is the thermal
diusivity of the layer and B
i
is determined by the boundary conditions. This approach
can be further simplied by introducing normalized variables. Let be the normalized time
dened as
=
t
t
f
, (4.15)
23
where t
f
is dened as the time it would take the temperature to diminish 99.3%, or 5 time
constants. Let be the normalized temperature dierence, dened as
=
T
surf
(0) T
surf
(t)
T
surf
(0) T
surf
(t
f
)
. (4.16)
Equation (4.14) can be normalized using the denitions of and as follows:
() = 1
j

i=1

i
e

i
, (4.17)
where
i
=
A
i
T
surf
(0)T
surf
(f)
and
i
=

i
t
f
. A least-squares method can be used to solve for
the
i
and
i
of each layer, both of which are normalized between 0 and 1. The synthetic
signal can be reconstructed using the tted coecients and (4.17).
In addition to performing a pure reconstruction, the synthetic signal can be modied.
For example, the time points used to reconstruct the signal can be altered to eectively
upsample, downsample, or interpolate the data [97].
4.2.6 Contrast Denitions
Four commonly used denitions of contrast are: absolute, running, normalized, and
standard contrast. Each of the techniques are outlined below and are summarized [30]. The
absolute contrast is dened as the excess temperature over a defect-free region at a given
time t and is given by
C
abs
(t) = T
defect
(t) T
defectFree
(t), (4.18)
where T is the temperature over a defect region and a defect-free region, respectively. This
increases the contrast and improves the visibility of the defective region over the defect-free
region.
The running contrast reduces the eects of dierences in surface emissivities and is
dened as
C
run
(t) =
C
abs
(t)
T
defectFree
(t)
. (4.19)
24
Note that if the contrast images are post-processed with the contrast stretching techniques
given previously, then the absolute contrast and running contrast are the same.
The normalized contrast can be computed with respect to the end of the thermal
process, at time t
end
, or the time of temperature max, t
max
(for pulsed thermography, this
is the rst frame). The normalized contrast is dened as
C
norm
(t) =
T
def
(t)
T
def
(t
max
)

T
defectFree
(t)
T
defectFree
(t
max
)
, (4.20)
where t
max
can be replaced with t
end
.
Finally, the standard contrast was developed to eliminate contributions from the sur-
rounding environment by subtracting pre-ash information given at time t
0
C
std
(t) =
T
defect
(t) T
defect
(t
0
)
T
defectFree
(t) T
defectFree
(t
0
)
. (4.21)
Each of these contrast images requires selection of a defect-free area.
4.2.7 Dierential Absolute Contrast (DAC)
The contrast methods described previously are greatly aected by non-uniform heating
of the surface and requires selection of a defect-free area. Dierential Absolute Contrast
(DAC) [37] removes the need for manual selection of a defect-free area and is therefore more
robust when non-uniform surface heating occurs. Let t

be the time at which the defect

begins to appear in the sequence and T(t) represent the frame at time t. Dene the
defect-free area as
T
snd
(t

) = T(t

). (4.22)
From (4.2), the value of
Q
e
can be solved to obtain
Q
e
=

T(t

). (4.23)
25
Using this result in (4.2), the ideal defect-free area can be found as
T
snd
(t) =
_
t

t
T(t

). (4.24)
By the denition of the defect-free region combined with the denition of absolute
contrast given in (4.18), the DAC image is given by
DAC(t) = T(t)
_
t

t
T(t

). (4.25)
Since the input pulse is not an ideal impulse, small dierences in pulse length can be
accounted for within dierent parts of the image. Dening t
e
to be the amount of error in
pulse length, a t can be made in the logarithmic domain to nd t
e
and compensate for the
error. The error-compensated DAC image is then found as
DAC(t t
e
) = T(t t
e
)
_
t

t
e
t t
e
T(t

t
e
), (4.26)
where the values of t
e
can vary over every pixel. The denition of the defect-free area can
also be used with the other denitions of contrast given in Section 4.2.6. A technique called
Interpolated Dierential Absolute Contrast (IDAC) has been developed to remove the need
for manual selection of time t

[40]. In addition, thermal quadrapole theory has been used

to extend the validity of DAC to later times [41, 42].
4.2.8 Derivative Images
A defect can theoretically be detected without the use of a reference region by evaluating
deviations from the ideal
1
2
slope seen in (4.11). To increase defect contrast and limit the
eects of blurring caused by lateral diusion, the derivatives of the analytical model for the
data are taken. The derivatives facilitate detection of an earlier time of maximum contrast,
thus reducing lateral diusion blurring. However, the diameter of the subsurface defect
must be greater than its depth beneath the surface for the lateral diusion to be eectively
ignored [51].
26
This algorithm begins by reconstructing the signal using the TSR technique described
in Section 4.2.5 to obtain an analytical expression for the data. The pixel time histories are
dierentiated using the expressions [52]
d ln(T
surf
(t))
d ln(t)
=
N

n=0
na
n
ln(t)
n1
, (4.27)
d
2
ln(T
surf
(t))
d ln(t)
2
=
N

n=0
n(n 1)a
n
ln(t)
n2
. (4.28)
The reconstructed signal and its time derivatives are transformed into the linear time
domain through (4.13) and by taking the exponent of the function as shown
dT
surf
(t)
dt
= exp
_
N

n=0
na
n
ln(t)
n1
_
, (4.29)
d
2
T
surf
(t)
dt
2
= exp
_
N

n=0
n(n 1)a
n
ln(t)
n2
_
. (4.30)
The resulting derivative time sequences can be outputted for analysis. Quantitative defect
depth analysis can be estimated as described by Omar et al. [45, 53, 54].
4.2.9 Principal Component Thermography (PCT)
Principal Component Thermography (PCT) [56, 57] uses singular value decomposition
(SVD) to reduce data to a compact statistical representation of the spatial and temporal
variations relating the contrasts associated with underlying material defects [56]. In ash
thermography data, a series of 2D image frames are stored sequentially, creating a 3D data
set. In order to perform PCT, a raster-like operation must be performed to create a 2D
representation of the 3D data. Given that the original data are loaded into an image cube
with dimensions N
x
, N
y
, and N
t
, where the N
x
and N
y
describe the pixel dimensions of
each frame and N
t
describes the number of frames. This image cube is then transformed
into a matrix A with dimensions MN
t
, where M = N
x
N
y
. The column vectors of M are
standardized to correct for individual detector pixel characteristics. This standardization
27
is achieved through

A(n, m) =
A(n, m)
n

n
, (4.31)
where

n
=
1
N
t
Nt

n=1
A(n, m), (4.32)

2
m
=
1
N
t
1
Nt

n=1
(A(n, m)
n
)
2
. (4.33)
Any M N matrix can be decomposed through Singular Value Decomposition (SVD)
into the following elements
A = UV
T
, (4.34)
where is a diagonal matrix containing the singular values of matrix A, and U and V
T
contain the left and right singular vectors of A. In this application, the matrix U contains
a set of orthogonal basis functions that describe the spatial variations within the data and
the matrix V
T
contains the corresponding characteristic time behavior which can be used
to estimate defect depths. By reversing the raster transformation applied to create A on
U, the empirical orthogonal functions (EOF) of the data are produced.
An analysis was performed showing that the rst two modes tend to contain 99% of
the variance within the data, although some leakage occurs in the higher order modes [57].
The rst mode describes a response similar to that of a uniform slab. However, the second
mode, characterizes a non-uniform eld created by material anomalies, and therefore has
been named the primary contrast mode or PCM. A drawback of the original PCT algorithm
is that it may enhance some defects at the cost of other defects [60]. The aw depth may be
estimated from the second principal component (PC), contained in the matrix V
T
, and by
knowing the thermal diusivities of the material using the technique developed in [56, 98].
4.2.10 Dynamic Thermal Tomography (DTT)
Two versions of the DTT algorithm were implemented, those referred to as classical
28
DTT and as reference-free DTT [5860]. The classical algorithm requires the selection of a
reference, defect-free region. The frames are rst normalized by dividing each frame with
the rst post-ash frame. The dierence between the time sequence of each pixel and the
reference signal is taken as
T(x, y, t) = T(x, y, t) T
ref
(t). (4.35)
Two images can be created from the new T signal: (1) an image of the maximum value of
T and (2) the time at which the maximum value occurs. These images are referred to as
a maxigram and a timegram, respectively. This creates a synthetic image that is sampled
at each pixels optimal time. The corresponding transit times in the timegram can be
used to create tomographic slices of the document.
The reference-free approach is similar to the classical approach, but removes the need
for manual selection of the defect-free region. Instead, dierent order polynomials are tted
to the normalized temperature response. The low-order polynomial will only reect general
behavior of the material, whereas higher orders will include the behavior of defects. In this
case, a third and a sixth order polynomial was used. The dierence for each pixel can be
found through
T(x, y, t) = T
h
(x, y, t) T
l
(x, y, t). (4.36)
The maxigram and timegram are obtained from T(x, y, t) and outputted. Defect depths
can be found using procedures described elsewhere [99].
4.2.11 Pulse Phase Thermography (PPT)
PPT is a combination of two forms of thermography, ash (pulse) thermography and
modulated thermography. In ash thermography a pulse of heat energy is deployed into the
target and the transient decay of the resulting surface temperature is analysed. Alternately,
in modulated thermography the target is subject to a sinusoidal temperature stimulation
in which standing thermal waves are created within the material. These standing thermal
29
waves are analyzed by their magnitude components and phase shifts with respect to the
reference modulation. The magnitude images are proportional to local optical and infrared
surface features; however, the phase shift images are relatively independent of these features.
As a result, the phase images can probe roughly twice the thickness given by the magnitude
image and are therefore the output of interest [46].
PPT uses the principle that a pulse of energy in the time domain contains all frequencies
in the frequency domain. Since the input pulse is not an ideal delta function, but rather a
rectangular pulse, the resulting frequencies are given by a sinc function [30]
F(f) = A
p
T
s
sinc(fT
s
), (4.37)
where f is the frequency variable, A
p
is the pulse amplitude, and T
s
is the sampling rate. In
eect, all frequencies are analyzed simultaneously in PPT rather than at a single frequency
as in modulated thermography [44]. In PPT, the discrete Fourier transform of each pixels
time series is computed using
F(u) =
1
N
N1

n=0
h(x)e
j2ux/N
= R(u) +jI(u), (4.38)
where R(u) and I(u) are the real and imaginary parts, respectively, of the transformed
sequence, F(u). The magnitude and phase responses are obtained from the transformed
data through
(u) = tan
1
_
I(u)
R(u)
_
, (4.39)
|F(u)| =
_
R(u)
2
+I(u)
2
, (4.40)
resulting in a series of magnitude and phase dierence output images.
As previously stated, the phase images are usually of interest due to their relative
independence of medium surface features. The resulting series of images correspond to
frequencies ranging from 0 to 1/t, where t is the time interval between images. The
30
lower the frequency, the deeper the image is able to probe. For the phase oset images, it
was found most useful if the maximum phase oset,
max
, was found for each pixel time
history and outputted into a single resulting image [44].
Another form of PPT is computed using the Wavelet Transform (WT) in place of
the Fourier Transform (FT). The advantage is that wavelets preserve time information of
the signal and are correlated to defect depth, allowing quantitative evaluations [47, 48].
Another technique uses the Hough Transform to retrieve the blind frequencies, which are
correlated with the defect depth [49]. It was also found that pre-processing the images with
the reconstruction technique given in TSR improved the depth resolution of PPT [50].
4.2.12 Correlation Images
The tting-correlation algorithm (FCA) [61] begins by reconstructing the signal using
the technique specied in TSR (see Section 4.2.5). The reconstructed signals are then
evaluated to see how closely they match either an ideal signal or a manually selected
defect-free signal. The ideal signal can be found using (4.5) or (4.24). Two methods are
used to evaluate the closeness of the t, the correlation coecient and the angle cosine.
The correlation coecient is calculated as
r =

n
i=1
(x
i
x)(y
i
y)
_

n
i=1
(x
i
x)
2
_

n
i=1
(y
i
y)
2
, (4.41)
and the angle cosine as
cos =

n
i=1
x
i
y
i
_

n
i=1
x
2
i
_

n
i=1
y
2
i
. (4.42)
The resulting correlation coecient image and the angle cosine image are output.
The correlated contrast technique [62] builds upon the FCA algorithm. Klein argues
that the chosen reference signal is irrelevant because dierences can be adjusted for by
changing the color map. In addition, it is noted that the dynamic range of the correlation
coecient image can be large, thus logarithmic color palettes are more appropriate than
liner color palettes for visualizing the images [62]. The correlated contrast implemented in
31
the IR View toolbox uses a fth order logarithmic root.
4.2.13 Transmission Line Matrix Fitting Algorithm (TLMFa)
The transmission line matrix is a numerical technique commonly used to model volt-
ages transmitted within transmission lines and has been extended as a method to model
forward diusion. Transmission lines can be modeled using the lossy wave equation, or the
telegraphers equation, as given

2
v = L
d
C
d

2
v
t
2
+R
d
C
d
v
t
, (4.43)
where v = v(x, y, z, t) and represents the voltage, R
d
, C
d
, and L
d
are the distributed param-
eters of a lumped transmission line and are the resistance, capacitances, and inductance.
Now to begin applying this to thermal diusion, interpret R
d
and C
d
as the thermal resis-
tance and capacitance per unit length, respectively. Also let the voltage be analogous to the
temperature v = u. By making appropriate space and time discretizations, the inductance
term will disappear reducing (4.43) to

2
u = R
d
C
d
u
t
. (4.44)
By dening =
1
R
d
C
d
, the equation above can be rewritten as

2
u =
1

u
t
, (4.45)
which, letting represent the thermal diusivity, is the thermal diusion equation describing
heat ow. Thus, given small enough space and time discretizations, thermal diusion can be
modeled using the telegraphers equation where the voltage is equivalent to the temperature
and the current is equivalent to the heat input. The only errors introduced are those
imposed by the space and time discretizations [66].
The TLM algorithm can be based o two models for a transmission line, the T-network
or -network models. These lead to two formulations of the TLM algorithm, referred to
32
as the link-line TLM node or link-resistor TLM node, respectively. Both methods achieve
the same results but have dierent design approaches. The link-line nodal arrangement
is derived in the following paragraphs. From the derivation above, the temperature is
analogous to the voltage and the heat energy is analogous to the current. These analogies
will be used for the rest of the derivation.
The transmission line can be seen as ideal segments of line with lumped resistance and
capacitance. An impulse entering this line will travel unimpeded for time
t
2
at which point
it encounters a discontinuity. Some of this initial pulse will transmit on and part will be
reected back. This is determined by the reection and transmission coecients which are
given by
=
R
R +Z
, (4.46)
=
Z
R +Z
. (4.47)
Assuming two input impulses are given approaching node x at time k. These pulses can
be represented as
i
k
V
L
(x) and
i
k
V
R
(x) (the i indicates that it is incident on node x). The
voltage at the node is given by the sum of the left and right bound pulses
k
(x) =
i
k
V
L
+
i
k
V
R
. (4.48)
The incident pulses are then scattered using the transmission and reection coecients
given above
s
k
V
L
=
i
k
V
L
+
i
k
V
R
, (4.49)
s
k
V
R
=
i
k
V
L
+
i
k
V
R
. (4.50)
The scattered pulse then travel along the line until they become incident on adjacent nodes
i
k+1
V
L
(x) =
s
k
V
R
(x 1), (4.51)
33
i
k+1
V
R
(x) =
i
k
V
L
(x + 1). (4.52)
The process of summation (4.48), scatter (4.50), and connect (4.52) are repeated to create
the link-line TLM algorithm.
Using a least squares tting method, such as the Levenberg-Marquardt Method, the
TLMFa algorithm can be applied to each individual pixels time response sequence and the
reection coecients can be determined. Since the reection coecients are dependent on
the thermal properties of the material, outputting images created from the each pixels
reection coecient for each node provides a characterization of a defect.
4.2.14 Hough Transform Thermography (PTHTa)
The Hough transform is a geometrical transform that is used in image processing to
nd geometrical structures within an image. The Hough transform is represented by
= xcos +y sin , (4.53)
where the points (x, y) are transformed into the Hough space (, ). The pulse thermography
Hough transform algorithm (PTHTa) [63, 64] uses the Hough transform to evaluate thermal
sequences removing the need for an operator by identifying pixels that follow the -
1
2
slope
shown in (4.11). Each point is transformed pixel-wise into sinusoidal curves in the Hough
space. The Hough space acts as an accumulator that sums the votes of all pixels in the
sequence providing an indication of the points that follow a -
1
2
slope. The points that
correspond to a -
1
2
lie in the 1.1071 rad column in Hough space [63]. By analyzing the
distribution of values in the 1.1071 rad column, defective areas can be separated from
non-defective areas. This is accomplished by performing a least-squares t of an exponential
function of the form
(x) = ae
(
xb
c
)
. (4.54)
It has been shown that parameters a and c are highly correlated to the defect depth and
34
images created from these parameters eectively enhance the contrast between the defect-
free and defect regions [63].
4.3 Developed Techniques
In an attempt to improve upon the existing techniques, the following techniques were
developed theoretically and algorithms developed to implement each. A description of each
technique follows.
4.3.1 Time-Dierence Contrast
Time dierencing allows for the subtraction of two frames that occur at dierent times
in order to analyze temperature changes between frames. This allows for the analysis of
relative changes in temperature between frames. Due to the decay of temperature between
time frames, it is assumed that for a defect-free image, the two frames would be related by
a multiplicative scale factor .
I
t
1
= I
t
2
, (4.55)
where I
t
1
and I
t
2
are the images at times t
1
and t
2
, respectively. It is assumed that image
I
t
1
is chosen such that the defect has not appeared yet, and could be either a pre-ash image
or an image immediately post-ash. Also, it is assumed that image I
t
2
is an image when
the defect has begun to appear. The scale factor can be determined using a minimum
mean square error (MMSE) method to solve
= min

||I
t
2
I
t
1
||. (4.56)
After determining , the two images are then dierenced
I
d
= I
t
2
I
t
1
, (4.57)
where I
d
is the dierence image or contrast mask. Since the MMSE technique will minimize
the average error of the images, I
d
will be positive for areas where defects appear and
35
slightly negative in areas where the defect does not appear. This method can be used to
dierence two specied images of interest. An automated method found to be eective is
to select the images to dierence based on the time relation
t
2
= t
1
+n T
s
, (4.58)
where n is an integer and T
s
is the sampling rate. A value of n = 5 was found to be sucient
in most applications.
4.3.2 Total Harmonic Distortion (THD)
Total harmonic distortion is an idea borrowed from microelectronics and quantizes the
amount of harmonic distortion, or noise, in a signal. THD can be dened as the ratio of
the sum of powers to the power of the fundamental frequency.
THD
P
=

n=2
P
n
P
1
=

n=2
V
2
n
V
2
1
, (4.59)
where P
n
and V
n
represents the power and voltage of the nth harmonic, respectively. An
alternative measure of harmonic distortion is given by amplitude ratio rather than the power
ratio and is dened by
THD
A
=
_

n=2
V
2
n
V
1
. (4.60)
To apply this to ash thermography, rst inspect the Laplace transform of an ideal
diusion as given by (4.4). The Laplace transform is
F(s) =
Q
e

(0.5)

s
. (4.61)
By dening C =
(0.5)
2

q
e
, (4.61) can be rewritten as
F(s) = C
1

s
. (4.62)
This is the frequency response of an ideal, defect free region. Treating thermal reections
36
due to defects back to the surface as noise, the amount of noise can be quantized using THD
resulting in a measurement of the degree in which a subsurface defect causes a deviation
from the defect-free response. From inspection of (4.62), the fundamental frequency is
chosen as the rst frequency o zero. Using both denitions of THD as given by (4.59) and
(4.60), the extent which a subsurface defect causes deviations from an ideal response can
be measured. Note that by dividing by the rst harmonic eectively normalizes the images
and thus is not aected by dierences in surface emissivities.
In addition to the THD, an alternate denition of signal-to-noise ratio can be calculated
as
SNR =

, (4.63)
where and are the mean and standard deviation in the frequency domain. The resulting
images tend to have a high spatial frequency and can be improved by using the image lters
described in Section 4.2.4 or by using a 2D Weiner lter.
4.3.3 Markov Error Contrast (MEC)
In MEC, each individual pixels time sequence is treated like a Markov chain, where
the next temperature in time is only a function of the current temperature. Recall (4.5)
describes the simplied response of a homogeneous slab given an impulse. Modifying this
equation, it can be shown that the next point is equal to
T
est
= T
current
_
t
c
t
c
+T
s
, (4.64)
where T
est
is the estimated next temperature state at time t
c
+T
s
, T
current
is the measured
current temperature state, t
c
is a scalar of the current time, and T
s
is the sampling rate.
The next temperature state is then estimated until every point has been evaluated and a
synthetic signal has been reconstructed. Note that every estimation is based o the current
measured value, not the estimated value for the current state. The error between the actual
37
and the estimated values for the entire sequence is found
E(t) = T(t) T
est
(t). (4.65)
The value of E(t) can be output for each pixel, creating a series of error images containing
the dierence between the actual data and the Markov estimation.
4.3.4 Time Constant Analysis (TCA)
The TCA algorithm is based o the apparent exponential decay of the thermal se-
quence. It performs a least-squares t of the temperature decay data to the model
T
surf
= ae
t

+c (4.66)
to describe the cooling of each pixel through time, where T
surf
represents the pixel temper-
ature at time t. The pixels rate of decay is described by the parameter , the pixels time
constant. An output image of each pixels time constant is created. Areas with defects will
have a slower time constant than sound regions.
4.3.5 Signal Detection and Matched Filtering (MF)
Matched ltering is used to increase visibility of defect areas while minimizing the
visibility of defect-free areas thus increasing the contrast of the material defects. Matched
ltering has been successfully used in MSI images to combine information across spectral
bands and enhance surface writings in ancient documents [100]. The method derived below
has been adapted and summarized [100, 101].
The MF technique is based o the following model
T
obs
= T
refl
+T
ideal
, (4.67)
where T
obs
, T
ideal
, and T
refl
is the observed temperature decay sequence, the ideal tempera-
ture decay sequence of a defect-free area, and the temperature prole caused by reections
38
from defects, respectively. The value is indicative of the strength of the reected signal.
To simplify notation on the derivations that follow, (4.68) is rewritten as
x = s +w, (4.68)
where x = T
obs
, s = T
refl
, and w = T
ideal
, where the bold font represents vectors. This
model requires a known temperature decay prole for a defect-free region. This can be
achieved by having an operator manually choose a defect free region. The matched ltering
techniques that follow multiply (4.68) by a vector q that maximizes the inuence from the
reected temperature and minimizes the defect-free response. The determination of q is
determined by the constraints
max
q
||q
T
s||
2
subject to min
q
||qw||
2
. (4.69)
All of the matched lter detectors accomplish the objective above, but use dierent
methods in determining q. The dierent matched lter detectors are outlined in the fol-
lowing paragraphs. For each of the lters, a known reection sequence, T
refl
, and the
ideal temperature sequence, T
ideal
, must be known. As previously stated, T
ideal
can be
determined by an operator choosing a background pixel or by using (4.5). Obtaining T
refl
can be a dicult, however the following method is proposed. First, the target pixels must
be selected. Then T
refl
is calculated by subtracting the ideal response from the observed
response leaving only the response from the reection
T
refl
= T
obs
T
ideal
. (4.70)
It was found that when an operator picked a signal, better signal-to-noise (SNR) ratio was
obtained if several signals were picked and the results were averaged. Several matched
ltering methods were implemented and the eectiveness of each evaluated.
The simplest matched lter technique is the Simple Matched Filter (SMF). The
SMF detector uses the known T
refl
sequence as q. This provides a correlation between the
39
known reection response and the observed temperature prole of each pixel, resulting in
a correlation image. The resulting SMF image is dened as
SMF = s
T
x
ij
, (4.71)
where the subscript i, j indicate the process is repeated over all the pixels time sequences
to create a single resulting image.
The Spectral Angle Map (SAM) detector builds upon the SMF detector through
normalizing by the vector magnitudes of the target and observed signals. The resulting
SAM image is given by
SAM =
s
T
x
ij

s
T
s
_
x
T
ij
x
ij
. (4.72)
The SAM image is the cosine of the angle between s and x. Note the SMF and SAM images
are similar in concept to the correlation techniques described in Section 4.2.12.
The Clutter Matched Filter (CMF) technique incorporates information arising
from the document structure by introducing the covariance matrix C. A derivation of
this lter using the probabilities of a defect-free and a defect containing region using the
Likelihood Ratio Test to develop a threshold on target presence is shown elsewhere [101].
The CMF image is given by
CMF = s
T
C
1
x
ij
, (4.73)
where C is the covariance matrix of the ideal diusion process
C =
1
m
m

i=1
ww
T
. (4.74)
The Adaptive Coherence Estimator (ACE) builds upon the CMF by normalizing by
the vector magnitudes, similar to how SAM builds upon SMF.
ACE =
s
T
C
1
x
ij

s
T
C
1
s
_
x
T
ij
C
1
x
ij
(4.75)
40
The Matched Subspace Filter (MSF) is used to detect subsurface defects that may
have dierent intensity reections or when detecting dierent subsurface defects but want
a single statistic to analyze the entire image. The matrix H is formed, where the columns
of H are the chosen target signals
H = [s
1
, s
2
, s
3
, ..., s
n
]. (4.76)
The measurement of energy of the observed signal projected into the target space, H can
be shown to be
MSF = x
T
H(H
T
H)H
T
x. (4.77)
The t-statistic and F statistic methods come from regression literature and the
derivations for both can be found elsewhere [101]. The t-statistic is simply the square root
of the F statistic. The equations for both are shown
tstat =
s
T
R
1
x
ij
_
x
T
ij
R
1
x
ij

2
(s
T
R
1
x
ij
)
2

d 1, (4.78)
Fstat =
(s
T
R
1
x
ij
)
2
x
T
ij
R
1
x
2
(s
T
R
1
x
ij
)
2

2
(d 1), (4.79)
where =
1

s
T
R
1
s
.
Other matched lter denitions implemented include: the Distance Matched Filter
(DMF) given by
DMF =
(
s

b
)
T
R
1
b
(x
b
)
(
s

b
)
T
R
1
b
(
s

b
)
, (4.80)
the Constrained Energy Minimization (CEM) given by
CEM =
s
T

x
1x
ij
s
T

1
x
s
, (4.81)
and the Minimum Variance Distortionless Response (MVDR) is given by
MV DR =
(s
x
)
T
R
1
x
(x
x
)
(s
x
)
T
R
1
x
(s
x
)
. (4.82)
41
In addition, other signal detection techniques were implemented such as RX Anomaly
Detection (RX)
RX = (x
b
)
T
R
1
b
(x
b
), (4.83)
the Quadratic Detector
QD = (x
b
)
T
R
1
b
(x
b
) (x
s
)
T
R
1
s
(x
s
), (4.84)
and Quadratic Discriminant Analysis (QDA)
QDA =
1
2
log |R
s
|
1
2
(x
ij

s
)
T
R
1
s
(x
ij

s
). (4.85)
Derivations for these techniques can be found elsewhere [100102]. Each of these tech-
niques were implemented, and evaluated for their eectiveness in processing ash thermog-
raphy data.
4.3.6 Convex Optimization Signal Detection Technique
In addition to the signal detection techniques described above, a technique using convex
optimization was attempted. Let every pixel, x
i
(t), be composed of an ideal base signal c(t)
and a sum of reection signals h(t) such that
x(t)
i
=
L

k=0
a
k
h(t k) +bc(t), (4.86)
where b is a scaling factor on c and a
k
is a sparse matrix of lter coecients. Re-writing
(4.86) in vector notation
x
i
= Ha
i
+e
i
, (4.87)
where e
i
is the error,
x
i
= [ c h(0) h(1) ... h(n) ], (4.88)
42
and H consists of a matrix of target vectors. The sparse matrix a
i
can be solved for by
min ||a
i
||
1
subject to ||Ha
i
x
i
||
2
, (4.89)
where is the allowed error. The values of a can be shown for every pixel sequence, creating
n coecient images. The values contained in a will indicate the strength of the target signal.
The diculty with this method is determining the target signals to ll the H matrix
with. Several options include manual selection of signals, randomly choosing signals, or
creating reection signals from nite element solutions (every unique structure would need
to be simulated). This technique is still under development and the results are not included
in this thesis.
43
Chapter 5
Analysis of Results
5.1 Processing Techniques Results
Each of the algorithms discussed in Chapter 4 were implemented and the eectiveness
of each was evaluated. Since it is dicult to dene general quantitative measures for image
quality, subjective measures of eectiveness were used with conclusions developed from the
experience of processing a variety of document structures presented. Representative images
were chosen to include in this report; however, the conclusions are generalized based upon
the experience of processing a variety of constructed data sets using various parameters.
Each of the algorithms studied were developed for, and perform best on, homogeneous
single-layer materials. Since ancient documents are often constructed with a variety of
materials and may have multiple layers, the evaluations were focused on the eectiveness
of processing complex structures.
The chapter is split into several parts. First, each technique is applied to a sample
data set and the imaging results showing the eectiveness in revealing subsurface inks
are presented. Next, techniques which prove applicable to document strikeouts and surface
writing enhancement are reviewed. Finally, comparisons are made between the eectiveness
of ash thermography and MSI for a variety of test documents. The results presented
throughout this chapter will be summarized in Chapter 6.
5.1.1 Subsurface Inks
Throughout this section, several images are presented demonstrating and comparing
the eectiveness of the processing routines. The constructed data set used for this technique
comparison is given in Figure 5.1. The example document was constructed by using an
44
adhesive to thermally bond the card stock sheet in Figure 5.1(a) to the card stock sheet
in Figure 5.1(b), with the surfaces containing the ink bonded together. Note that Figure
5.1(a) has been mirrored in order to reect the direction the ink would face when the layers
are glued together. Each processing technique will be applied to the ash thermography
data of Test Document 1 and the results discussed.
Noise Reduction
Several techniques were used in reducing noise within the images. A median lter
was used to remove salt-and-pepper noise. A low-pass Gaussian lter was used to remove
high spatial frequency content that is often a result of noise. Finally, temporal noise was
reduced using either the thermal signal reconstruction (TSR) or the layered reconstruction
(LR) synthetic signal techniques.
It was found that using a 33 lter window was sucient for both the median and the
Gaussian lter approaches. In addition, a variance of approximately 0.8 was found to be
eective for the Gaussian lter. Finally, the thermal signal reconstruction was found to be
preferred over the layered reconstruction.
Although both generally reconstructed the signal adequately, the TSR method was
found to be computationally faster (see Table 5.1) and generally reconstructed the signal as
good or better than the LR method. For the TSR method, using a sixth order polynomial t
appeared to give sucient results. In LR, a minimum of three layers were usually required
(a) Top Side (Mirrored) (b) Bottom Side
Fig. 5.1: Test document 1.
45
to give an accurate reconstruction; however, using more layers increased the accuracy (and
computation time) of the reconstructed signal. Figure 5.2 shows the results of the noise
reduction process. Note that these images have been normalized to increase contrast.
The images presented throughout this chapter and in Figure 5.2, unless specically
stated otherwise, were pre-processed using the median lter, followed by the low-pass Gaus-
sian lter, followed by the TSR reconstruction. This approach appeared to be the more
eective order for application of the noise reduction techniques.
Pseudo-Color Images
Figure 5.3 present the use of pseudo-color images on a sample raw frame. In Figure
5.3(a), very little ink information can be detected. However, using pseudo-color, as shown in
Figure 5.3(b), the aw information can be faintly observed. Consistent with the literature
[30], it was found that using a continuous color map was most eective for this application.
Pseudo-color images can be used to evaluate the unprocessed frames (as shown in Figure
5.3) or to better visualize the processed data. It was found useful to evaluate the thermal
images in both pseudo-color and grayscale.
Contrast Stretching and Enhancement
The techniques discussed in this section, such as image normalization and histogram
(a) Raw (b) Thermal Signal Reconstruc-
tion (TSR)
(c) Layered Reconstruction
(LR)
Fig. 5.2: Noise reduced images at t = 228.8 ms.
46
Table 5.1: Signal reconstruction wall times.
Algorithm Wall Time (sec)
TSR 19.12
LR 3 900.3
LR 4 1638
(a) Original (b) Pseudo-Color
Fig. 5.3: Pseudo-color image at t = 228.8 ms.
equalization, can be used to successfully process the raw images. However, the thermal-
based processing algorithms discussed in the following sections usually provide better defect
visualization. The advantage of contrast stretching techniques is highest when used as a
post-processor to the thermal techniques discussed later.
Figure 5.4 compares the results of the same frame contrast enhanced using normaliza-
tion, normalization with end saturation, and histogram equalization. An increased amount
of subsurface ink is observed in Figure 5.4 compared with the non-contrast enhanced image
shown in Figure 5.3(a). Figure 5.4(c) demonstrates that histogram equalization can be a
very useful image enhancement technique; unfortunately, undesirable image deformations
may occur.
Contrast Denitions and Dierential Absolute Contrast
The Dierential Absolute Contrast (DAC) technique was found to be a very eective
contrast technique, depending on dierential frame, t

, selected. The advantage of this tech-

nique is that by choosing t

values relating to defects at dierent depths, specic defects

47
(a) Normalization (b) Normalization with End
Saturation
(c) Histogram Equalization
Fig. 5.4: Contrast stretching at t = 228.8 ms.
can be emphasized in contrast with others. This is advantageous when applied to docu-
ments with complex structures because the writing layer can be selected and emphasized
as illustrated in Figure 5.5. Figure 5.5(a) shows an eective choice of sample frame and
dierential contrast image choice to reveal the top layer of ink. On the other hand Figure
5.5(b) better reveals the bottom layer of ink.
The images from the other contrast denitions are presented in Figure 5.6. It is ap-
parent that the DAC technique outperforms the standard contrast denitions. Note that
using the rst post-ash frame, instead of the nal sample frame, resulted in more eective
contrast when using normalized contrast approach.
(a) Frame at t = 46 ms shown
using dierential image at t

=
286 ms
(b) Frame at t = 515 ms shown
using dierential image at t

=
572 ms
Fig. 5.5: Dierential absolute contrast (DAC) images.
48
(a) Absolute Contrast (b) Standard Contrast
(c) Running Contrast (d) Normalized Contrast
Fig. 5.6: Contrast denitions.
Thermal Signal Reconstruction (TSR) and Derivative Images
The signal reconstruction and noise reduction of using the TSR technique proved to
be benecial in processing. The derivative images were found to be one of the techniques
which provided consistently good defect contrast enhancement. The technique was also
able to reveal some of the deep-seated defects. Figure 5.7 presents eective rst and second
derivative images.
Principal Component Thermography (PCT)
For single-layer homogeneous materials, PCT is eective in revealing a subsurface layer
of ink. For great structural complexity of documents, the PCT algorithm tends to enhance
some ink areas at the cost of others. The result was that obtaining good results was incon-
sistent in revealing subsurface ink. However, the PCT technique is an eective technique
as illustrated in Figure 5.8. It was found useful to inspect only the rst ve to ten principal
component images.
49
(a) First Derivative (b) Second Derivative
Fig. 5.7: Derivative images.
Fig. 5.8: Principal component thermography (PCT) images.
Dynamic Thermal Tomography (DTT)
The DTT approach proved to be mediocre in revealing subsurface inks. The DTT
algorithm was able to reveal the ink of the top layer, but was unable to reveal the deep-
seated layer of ink as shown in Figure 5.9. In addition, it was found that the reference-
free technique is preferred because it enhances the ink as well or better than the classical
technique without the need for manually selecting a defect-free region.
Pulse Phase Thermography (PPT)
The PPT phase images consistently revealed the subsurface defects. Due to the in-
creased independence to surface features, the PPT phase images also were found to be
eective in revealing the subsurface writing for complex document structures. As expected,
the low frequency phase images reveal deep-layered defects better than do the high frequency
phase images. Therefore, PPT is an essential processing tool for analyzing documents. Fig-
50
(a) Classical Maxigram (b) Reference-Free Maxigram
Fig. 5.9: Dynamic thermal tomography (DTT) images.
ure 5.10 presents the PPT phase image corresponding to 1.2 Hz. However, PPT algorithm
still is unable to reveal the bottom ink layer eectively in this document.
Correlation Images
The FCA algorithm performed moderately well in each test. In general, techniques
such as PPT, DAC, THD, MEC, or derivative images outperformed the FCA algorithm or
the correlated contrast image. The correlation images are presented in Figure 5.11.
Transmission Line Matrix Fitting Algorithm (TLMFa)
The transmission line matrix model tting algorithm (TLMFa) proved to be compu-
tationally intensive without providing signicant improvements in subsurface ink enhance-
ment. Therefore, the TLMFa algorithm is not recommended as a document thermography
algorithm. Only a simplied, 1D version of the algorithm was implemented. Implementing
Fig. 5.10: Pulse phase thermography (PPT) image at 1.2 Hz.
51
(a) Angle Cosine (b) Correlation Coecient (c) Correlated Contrast
Fig. 5.11: Correlation images.
a higher dimensional transmission line matrix would further increase the computational
expense, but could possibly provide better detail and be of interest for further research. A
detailed reection coecient image is shown in Figure 5.12. It was found that at least 25
transmission line nodes are required to create an eective tting.
Hough Transform Thermography (PTHTa)
The PTHTa algorithm was found to be useful in detecting the subsurface inks although
it was computationally intensive. The images corresponding to the coecients of the expo-
nential t are presented in Figure 5.13. The B coecient image shown in Figure 5.13(b)
reveals the deepest layer of subsurface inks. The PTHTa technique is an eective tech-
nique; however, due to its computationally intensive nature use of alternate techniques is
suggested.
Fig. 5.12: Transmission line matrix tting algorithm (TLMFa) image.
52
(a) A Coecient Image (b) B Coecient Image (c) C Coecient Image
Fig. 5.13: Hough transform coecient images.
Time-Dierence Contrast
The time-dierence contrast technique was found to be among the more eective tech-
niques. This technique works eectively when dierencing each frame from the fth previous
frame. Results from this technique are given in Figure 5.14, where Figure 5.14(a) shows the
top layer of ink eectively and Figure 5.14(b) reveals the bottom layer of ink eectively.
Total Harmonic Distortion (THD)
The THD technique proved to be an eective technique as illustrated in Figure 5.15.
The THD technique is able to reveal the deep-seated layers of ink as eectively or more
so than the other methods used. Both the THDa and THDp versions yield similar results
and are able to reveal the deep-seated inks better than the SNR image. An advantage that
THD has over the contrast images is that it is able to reveal the top and bottom layers
(a) t = 114 ms (b) t = 515 ms
Fig. 5.14: Time-dierence contrast images.
53
of ink eectively in a single image. The disadvantage of THD is that the visualization of
the inks are aected by lateral diusion more than in techniques such as PPT, causing an
increase in blurriness of the detected subsurface inks.
Markov Error Contrast (MEC)
The Markov Error Contrast (MEC) technique was able to eectively reveal both surface
and deep-seated layers of writing as well or better than alternate techniques as can be
seen in Figure 5.16. The MEC technique provides a new useful tool for processing ash
thermography data. The MEC technique has an advantage over the DAC technique, both
of which give similar results, in that the MEC technique does not require selection of a
dierential frame.
Time Constant Analysis (TCA)
The time constant analysis (TCA) technique performed eectively on single-layer doc-
ument structures. It proved to be an eective technique for analyzing ash thermography
video streams as can be seen from the results shown in Figure 5.17. Because of the compu-
tational expense of TCA, other techniques with similar eectiveness, such as PPT, MEC,
or DAC, are often preferable.
(a) THDa (b) THDp (c) SNR
Fig. 5.15: Total harmonic distortion (THD) images.
54
(a) t = 114 ms (b) t = 572 ms
Fig. 5.16: Markov error contrast (MEC) images.
Fig. 5.17: Time constant analysis (TCA) image.
55
Signal Detection and Matched Filtering
Four of the matched lter methods consistently performed well in the tests: the Spectral
Angle Map (SAM), Adaptive Coherence Estimate (ACE), the t-statistic, and the F-statistic.
The results for using each of the four lters are presented in Figure 5.18. These techniques
are able to eectively reveal both layers of ink in a single image. More eective results are
obtained when the deep-seated ink layers are chosen as the target pixels.
5.1.2 Document Strikeouts
The following section presents the results for revealing strikeouts contained in surface
writings of a document. Strikeouts refer to areas where two layers of the same ink overlap
as a result of textual changes. It is important to note that each image was pre-processed
using the noise reduction technique previously described and that the results were manually
contrast enhanced.
The document sample shown in Figure 5.19 was constructed to evaluate the eective-
ness of imaging document strikeouts. The sample data set was constructed using iron gall
ink on a white card stock substrate. It consists of an X shaped ink under-layer with hor-
izontal stripes overlaying the X. The algorithms that performed eectively in revealing
strikeouts are discussed and their results demonstrated in this section.
Contrast Denitions and Dierential Absolute Contrast (DAC)
Three of the contrast denitions were found useful for evaluating strikeouts: the ab-
solute contrast, running contrast, and normalized contrast. The most eective results for
each contrast denition can be seen in Figure 5.20. The absolute and running contrast
approaches give very similar images in which half of the general shape of the X can be
seen. However, the normalized contrast gives the best results of the three methods since the
entire ink structure can be seen. Notice in the normalized contrast shown in Figure 5.20(c),
the areas appear brighter where the two layers of writing overlap. In the upper right hand
corner, the rectangular shape of the underlying X can be seen clearly. In addition, the
dierential absolute contrast also performed well as illustrated in Figure 5.21.
56
(a) SAM (b) ACE
(c) f-statistic (d) t-statistic
Fig. 5.18: Matched lter images.
Fig. 5.19: Strikeouts test set.
57
(a) Absolute Contrast (b) Running Contrast (c) Normalized Contrast
Fig. 5.20: Contrast denition strikeout images.
Fig. 5.21: DAC strikeout image with t

= 11.5 ms and t = 57.5 ms.

58
Dynamic Thermal Tomography (DTT)
For detecting strikeouts, the classical DTT algorithm signicantly outperformed the
reference-free DTT algorithm. This is the opposite conclusion that was drawn when detect-
ing subsurface inks. The results of the DTT algorithm can be seen in Figure 5.22.
Markov Error Contrast (MEC)
The Markov Error Contrast (MEC) algorithm proved to be useful in the detection of
document strikeouts. The results from MEC are shown in Figure 5.23.
Principal Component Thermography (PCT)
For the PCT algorithmic approach, it was found that decreasing the total time for
sampling improved the PCT results. Figure 5.24(a) shows the results for a sampling time
of 2.2 s, where Figures 5.24(b) and 5.24(c) show the results for the reduced total sampling
time of 172.4 ms. For PCT, reducing the number of frames increased the detectibility of
the strikeouts.
Pulse Phase Thermography (PPT)
The PPT algorithmic approach, using the amplitude images rather than the phase
images was found to be benecial in analyzing strikeouts. In addition, it was found higher
frequencies were more useful for revealing strikeouts. The result of PPT corresponding to
43.7 Hz is shown in Figure 5.25.
Fig. 5.22: DTT classical maxigram.
59
Fig. 5.23: Markov error contrast (MEC) strikeout image.
(a) PCM (sampled For 2.2 s) (b) PCM (sampled For 172.4
ms)
(c) First principal component
(sampled For 172.4 ms)
Fig. 5.24: Principal component thermography (PCT) strikeout images.
Fig. 5.25: Pulse phase thermography (PPT) amplitude strikeout image.
60
Transmission Line Matrix Fitting Algorithm (TLMFa)
The results from using the transmission line matrix tting algorithm (TLMFa) are
shown in Figure 5.26. Despite being able to eectively reveal the strikeouts, the extreme
computational expense of this technique makes other methods preferred.
Time Constant Analysis (TCA)
The results of the TCA approach can be signicantly improved for document strikeouts
using a shortened total sampling length. The images in Figure 5.27 were obtained using a
total sampling time of 172.4 ms.
Time-Dierence Contrast
The results of the time-dierencing contrast approach are given in Figure 5.28. The
time-dierence contrast images are eectively able to reveal the strikeouts.
Derivative Images
It was found that using a shortened total sampling time improved the results of the
TSR images for document strikeouts. The rst and second derivative images can be seen
in Figure 5.29.
Fig. 5.26: Transmission line matrix tting algorithm (TLMFa) strikeout image.
61
(a) Time Constant (b) Amplitudes
Fig. 5.27: Time constant analysis (TCA) strikeout image.
(a) Frame at 69 ms (b) Frame at 287.4 ms
Fig. 5.28: Time-dierence contrast strikeout images.
(a) First Derivative (b) Second Derivative
Fig. 5.29: Strikeout derivative images
62
5.2 Comparison of Algorithm Wall Times
A study was performed comparing processing wall times of each algorithm. The wall
time is a measure of the amount of chronological time it takes for a process to complete.
The tests were performed on a Intel R Core
TM
2 Quad 2.4Ghz CPU Q6600 with 4 GB of
memory. The data set used to test the algorithm wall times consisted of 189 frames with
each individual frame having dimensions of 72 by 81 pixels. The code was implemented in
Matlab with four parallel working nodes. Figure 5.30 shows a bar graph of the wall times
created from the data in Table 5.2. The algorithm numbers in Figure 5.30 correspond to
the algorithms listed in Table 5.2. Due to their signicant computational cost and lack of
signicant improvements, the TLMFa, PTHTa, and TCA algorithms are not recommended.
The contrast images tend to perform well and have minimal computational cost; as such,
they are a recommended starting point when processing ash thermography data.
5.3 Comparison with Multi-Spectral Imaging (MSI)
A comparative study was conducted to determine the eectiveness of MSI versus ash
thermography techniques. Each tested document type is discussed in the following sections.
5.3.1 Subsurface Inks
Several examples of subsurface inks using a variety of covering layers were tested.
Figure 5.31 shows a comparison of MSI images with the ash thermography images for
a textured card stock layer. For depth penetration in MSI, the 950 nm image proved to
give the best contrast as shown in Figure 5.31(a). Figures 5.31(b), 5.39(c), 5.39(d) show
the processed ash thermography images for comparison. Although the MSI approach is
able to faintly reveal the subsurface writing, the ash thermography method signicantly
outperforms the MSI image method in enhancing the contrast of the subsurface inks. In
addition, the ash thermography data set can reveal relative depth information for the inks,
which MSI is unable to accomplish. Each image in Figure 5.31 have been contrast stretched
using image normalization.
Additional comparisons of MSI results versus pulsed thermography results for each of
63
Fig. 5.30: Algorithm wall times.
Table 5.2: Algorithm wall times.
Number Algorithm Average Wall Time (sec)
1 DAC 0.03191
2 Absolute Contrast 0.03244
3 Normalized Contrast 0.03979
4 Standard Contrast 0.04003
5 Running Contrast 0.04579
6 MEC 0.1597
7 THD 0.3365
8 PPT 0.3465
9 PCT 0.7850
10 MF 1.075
11 TDC 1.357
12 DTT (Classical) 2.948
13 TSR 4.037
14 DTT (Reference-free) 6.843
15 Correlation Images 10.76
16 TCA 99.96
17 PTHTa 125.9
18 TLMFa 4044
64
(a) MSI 950nm (b) PT MEC (inverted)
(c) PT Time-Dierence (in-
verted)
(d) PT THDp
Fig. 5.31: MSI vs. FT strikeout results.
the subsurface document classes are shown in Figures 5.32-5.39. Three inks were tested:
(1) ball point (upper right), (2) carbon (upper right and lower left), and (3) iron gall (lower
right). Figure 5.32 shows the under-writing for two test documents covered with a sheet of
white card stock. Figure 5.33 shows the ash thermography results for the ink underwriting.
Unfortunately, an MSI comparison is not available for this data set. However, Figure 5.34
compares the detection of an Edmunds Optics Resolution Chart placed underneath a white
card stock layer. It can be seen that the ash thermography images are better able to reveal
the optics card.
The papyrus test document is shown in Figure 5.35. For the papyrus covering layer,
shown in Figure 5.36, the MSI image provides a cleaner result than the PT images. This is
expected for thin layers, such as this particular sheet of papyrus.
Figure 5.37 shows the paint test document and the imaging results are presented in
65
(a) Ink Layer (b) Optics Card
Fig. 5.32: Cardstock test document under-layers.
(a) PT MEC (inverted) (b) PT PCT (c) PT MF-ACE (inverted)
Fig. 5.33: MSI vs. PT cardstock results.
(a) MSI 950nm (b) PT MEC (inverted)
(c) PT TDC (inverted) (d) PT PPT
Fig. 5.34: MSI vs. FT cardstock 2 results.
66
(a) Ink Layer (b) Papyrus Covering
Fig. 5.35: Papyrus test document.
(a) MSI 950nm (b) PT PCT
(c) PT TDC (inverted) (d) PT PPT
Fig. 5.36: MSI vs. FT papyrus results.
67
Figure 5.38. Both the MSI and the PT methods are able to reveal the majority of the ink
underneath the painted layer. However, the PT techniques are able to reveal the writings
beneath the black paint which the MSI fails to reveal. This indicates that PT can be used to
reveal writings underneath a layer that has similar spectral properties but dierent thermal
properties.
Figure 5.39 demonstrates the ability of ash thermography to reveal subsurface inks
beneath a layer of gesso. In this case, the MSI technique performs poorly. The results
presented in this section demonstrate the ability of ash thermography to image subsurface
inks beneath a variety of covering materials.
5.3.2 Document Strikeouts
For document strikeouts, MSI fails completely to reveal the under-layer of writing as
shown in Figure 5.40(a). However, Figures 5.40(b) and 5.40(c) prove the ability of ash
thermography to detect document strikeouts. The ability of ash thermography to detect
document strikeouts is a signicant result that can allow for the evaluation of documents
that have previously been unable to analyze eectively.
5.3.3 Surface Writing Enhancement
The purpose of the test document shown in Figure 5.41(a) is to evaluate the ability
of ash thermography to enhance surface inks that have poor spectral contrast. Figures
5.41(b) and 5.41(c) show the results of two dierent wavelength MSI images. In the 950 nm
(a) Ink Layer (b) Paint Covering
Fig. 5.37: Paint test document.
68
(a) MSI 950nm (b) FT MEC
(c) FT THDp (d) FT PPT
Fig. 5.38: MSI vs. FT painted images.
(a) MSI 400nm (b) FT DAC
(c) FT PPT (d) FT Time-Dierence Con-
trast
Fig. 5.39: MSI vs. FT gesso images.
69
(a) MSI 950nm (b) PT MEC (inverted) (c) PT 2nd Derivative
Fig. 5.40: MSI vs. FT strikeout results.
image, the surface inks have poor contrast; however, in the 600 nm they have signicant
contrast. The two ink writings can also be seen in the ash thermography images shown in
Figures 5.41(d)-5.41(f). Although both the red and blue inks can be seen in both MSI and
FT images, the important dierence is the MSI images are comparing dierences in optical
properties whereas the FT images are comparing the dierences in thermal properties.
This indicates that ash thermography can be used to enhance the writings of surfaces
that contain poor spectral contrast but have a large dierences in thermal properties of
materials.
5.4 Feasibility of Application to Archaeological Artifacts
The methodology presented in this thesis was validated by applying the developed ash
thermography techniques to a piece of Egyptian cartonnage and the Codex Selden.
5.4.1 Egyptian Cartonnage
The cartonnage is an ancient document with a complex structure, containing overlap-
ping areas of ink, papyrus, gesso, and paint. Further imaging challenges arise from the
signicant material inhomogeneities and internal aws contained within the cartonnage.
Figures 3.2 and 5.42 provides comparisons between standard imaging, MSI techniques, and
ash thermography techniques.
70
(a) Test Document (b) MSI 950nm (c) MSI 600nm
(d) PT PPT (e) PT 2nd Derivative (f) PT MF-ACE
Fig. 5.41: MSI vs. FT red surface results.
71
From these gures, several conclusions may be reached. First, both the MSI images
and the ash thermography images are comparable in their ability to signicantly enhance
the cartonnage surface writing. The important dierence between the two methods is the
technique used to enhance the surface writing. MSI enhances this text through dierences
in the spectral characteristics of ink and the papyrus background. On the other hand, ash
thermography enhances the surface writing through dierences in the thermal properties of
the ink and the papyrus background.
To illustrate that the ink is not enhanced due to spectral characteristics in ash ther-
mography, a pre-ash cartonnage image is shown in Figure 5.43. In the pre-ash image, the
surface ink cannot be observed, indicating that dierences in spectral characteristics have
a insignicant contribution to the visualization of the ink. Thus text on ancient documents
that previously were not eectively imaged using MSI because of poor spectral contrast be-
tween inks and document backings, can now be imaged using ash thermography to reveal
these obscure writings.
A second important conclusion from this study is the improvement ash thermography
provides over MSI imaging when applied to ink just below the document structure. This
is illustrated in Figure 5.42(c). A red arrow points to a subsurface ink line that is not
observed in Figure 5.42(a). This indicates that ash thermography successfully revealed
shallow subsurface ink that MSI could not show.
5.4.2 Codex Selden
The Codex Selden is a Mesoamerican manuscript book that is a recognized palimpsest
[103105]. The original pictograph drawings were covered with a gesso layer, and a new
set of pictographs were drawn on the surface. The Codex Selden presents several unique
imaging challenges including: (1) an unknown thickness of the gesso layer, (2) an unknown
extent to which the original writing has been removed, and (3) internal aws within the
gesso that may obscure detectability of under-inks.
The Codex Selden data was collected using a ThermaCAM Phoenix 600 camera oper-
ating in the 1.5-5.0 m range with a data acquistion rate of 67 frames/second. Two Sunpak
72
(a) MSI 950nm
(b) FT PPT Amplitude Image (Inverted)
(c) FT Standard Contrast
Fig. 5.42: MSI vs. FT cartonnage images.
73
Fig. 5.43: Pre-ash FT cartonnage image.
Pro System 622 ash units provided the impulse energy [106]. Imaging access to the Codex
Selden was limited and the images presented were obtained early in this research before the
most e cient ash thermography parameters had been established.
The results of imaging three sections of the Codex Selden are presented in Figures 5.44,
5.45, and 5.46. In Figure 5.46(b), the vertical and horizontal line discontinuities within the
image are an artifact of combining several images. Deeper document structural defects can
be detected using ash thermography techniques over MSI as illustrated by the examples
given. It is too premature to conclude if any of these areas are potentially subsurface inks.
In addition, ash thermography enhances the thin surface inks, demonstrating that ancient
inks as well as modern ink can be enhanced using ash thermography techniques. It is
reasonable to assume su cient thermal bonding between the gesso and the subsurface inks.
Re-obtaining the data set with larger amplitude excitation pulse is recommended.
74
(a) Visible Image (b) MSI 1000 nm
(c) FT PPT (d) FT PCT
Fig. 5.44: MSI vs. FT Codex Selden images data set 1.
75
(a) Visible Image (b) MSI 750 nm
(c) FT PPT (d) FT MEC
Fig. 5.45: MSI vs. FT Codex Selden images data set 2.
76
(a) Visible Image (b) MSI 950nm
(c) FT PPT (d) FT MF t-statistic
Fig. 5.46: MSI vs. FT Codex Selden images data set 3.
77
Chapter 6
Conclusions and Recommendations
The conclusions developed from the results presented in Chapter 5 are summarized
in this chapter. The ash thermography parameters for eective ash thermography data
acquisition are outlined. A summary of methods for pre-processing, processing, and post-
processing the data to increase ink visualization are given. The ability of ash thermography
to image a variety of materials are discussed. Finally, this thesis concludes with recommen-
dations for future research.
6.1 Thermography Data Acquisition Parameters
Several ash thermography parameters were evaluated during the acquisition of the
data sets. Recommended values are discussed below.
Pulse Amplitude Four synchronized ash units were sucient to reveal subsurface layers
of ink (after processing) through paper, card stock, papyrus, and thin gesso layers.
Large pulse amplitudes are desirable, but care must be taken to reduce the residual
heating from the ash bulb which may occur. For document strikeouts or enhancing
surface writing, a large pulse amplitude is unnecessary.
Pulse Duration Residual heating from ash pulse tails were not found to have a strong
negative eect on the detection process for deep-seated subsurface writings, provided
the tail decayed relatively quickly and absorbed uniformly into the document. For
document strikeouts and surface ink enhancement, a minimal input pulse length that
closely approximates an impulse is desired.
Sampling Frequency A sampling frequency of 87 Hz was found adequate for subsurface
defects; however, a faster sampling rate is recommended for document strikeouts and
surface ink enhancement.
78
Lens Filters The 3.42 4.05 m lter greatly outperformed the 2.65 3.24 m lter.
6.2 Pre-Processing
The following techniques were found useful for pre-processing ash thermography data
when applied in the following order:
1. Subtraction of a pre-ash image to each frame in order to reduce eects of non-uniform
surface emissivities;
2. A 33 pixel median lter applied to each frame for removing outliers caused by dead
pixels;
3. A 33 Gaussian low-pass lter, with a variance parameter of 0.8, applied to reduce
spatial image noise in each frame;
4. Thermal signal reconstruction (TSR), using a sixth order polynomial tting, for re-
ducing temporal noise.
6.3 Processing
Eective processing techniques for each document class are presented. Table 6.1 out-
lines a rating system used in classifying the algorithms. Table 6.2 evaluates the pre-existing
and developed algorithm techniques used for processing ash thermography data of doc-
uments. Note that due to the similarities between processing strikeouts and surface ink
enhancement, the ratings for strikeouts also apply to surface writing enhancement.
6.4 Post-Processing
Essential post-processing methods include the following contrast stretching techniques:
1. Manual histogram manipulation,
2. Image normalization,
3. Image normalization with ends saturation,
79
Table 6.1: Algorithm rating system.
Value Denition Explanation
3 Highly Recommended
Consistently gives great results with minimal
computational expense.
2 Recommended
Performs well but inconsistently. May only
work for simple document structures or com-
putational expense is moderate.
1 Moderately Recommended
Performs well occasionally or has a signicant
computation expense.
0 Not Recommended
Either fails to perform well or computational
expense is prohibitive.
4. Pseudo-color images.
Other techniques which may be useful, depending on the situation, include:
1. Histogram equalization,
2. Image lter (median or Guassian),
3. Absolute and running contrast denitions.
6.5 Materials
This research demonstrated that ash thermography can be used to detect three types
of inks: carbon-based, iron gall, and modern ball-point. These inks were detected beneath
several covering layers including card stock, papyrus, paint, a gesso mix, and the ink itself.
6.6 Signicant Contributions
This research resulted in the following signicant contributions.
1. A comprehensive analysis of existing pulsed thermography processing techniques ap-
plied to documents, the magnitude of which does not exist in the literature. Pre-
existing techniques investigated include pseudo-color images, contrast stretching, syn-
thetic signal reconstruction, median and Gaussian low-pass lters, four contrast de-
nitions, dierential absolute contrast (DAC), derivative images, principal component
80
thermography (PCT), classical and reference-free implementations of dynamic thermal
tomography (DTT), pulse phase thermography (PPT), correlation images, transmis-
sion line matrix tting algorithm (TLMFa), and pulsed thermography Hough trans-
form algorithm (PTHTa).
2. New pulsed thermography processing techniques that improve upon the results of the
existing techniques were developed. These techniques include Markov error contrast
(MEC), time-dierence contrast, total harmonic distortion (THD), time constant anal-
ysis (TCA), and four matched lter denitions including spectral angle map (SAM),
adaptive coherence estimator (ACE), t-statistic, and F-statistic.
3. A proof-of-concept for detecting subsurface ink writings in documents.
4. Verication of the capability of pulsed thermography techniques to detect document
strikeouts.
5. Demonstrates the ability to enhance surface writings based on dierences in thermal
characteristics when spectral characteristics do not dier signicantly.
6. Demonstrates that pulsed thermography signicantly improves upon multi-spectral
imaging for subsurface and surface writing enhancement.
7. Provides an evaluation of ash thermography parameters for the most eective doc-
ument imaging.
8. Validates the ndings presented through application of ash thermography techniques
to a piece of Egyptian cartonnage and the Codex Selden.
9. A Pulsed Thermography Toolbox was developed providing a graphical user interface in
MATLAB for the straight-forward application of the presented processing techniques
to ash thermography data.
81
6.7 Recommendations for Future Research
There are two principal directions suggested for future research. The rst recommen-
dation is the development of advanced thermal models that account for the uneven, multiple
layers which are often found in ancient documents. From these advanced thermal models,
processing techniques with greater eectiveness can be developed specic to this applica-
tion. The second recommendation is research into alternate heating methods to facilitate
for greater excitation energy. Methods of future research could explore step heating, ther-
mal transfer heating, or modulated thermography techniques. For document strikeouts,
it is recommended that the experiments be repeated using a thermal camera capable of
signicantly faster frame rates. Investigation of techniques which account for the lateral
diusion through the ink would also be benecial.
82
T
a
b
l
e
6
.
2
:
P
r
o
c
e
s
s
i
n
g
a
l
g
o
r
i
t
h
m
s
s
u
m
m
a
r
i
z
e
d
.
P
r
o
c
e
s
s
i
n
g
T
e
c
h
n
i
q
u
e
S
u
b
s
u
r
f
a
c
e
R
a
t
i
n
g
S
t
r
i
k
e
o
u
t
R
a
t
i
n
g
W
a
l
l
T
i
m
e
(
S
e
c
)
C
o
m
m
e
n
t
s
C
o
n
t
r
a
s
t
D
e

n
i
t
i
o
n
s
1
1
0
.
0
3
9
5
1
C
a
n
b
e
u
s
e
d
a
s
a
p
o
s
t
-
p
r
o
c
e
s
s
o
r
.
(
4
.
2
.
6
)
D
i

e
r
e
n
t
i
a
l
A
b
s
o
l
u
t
e
C
o
n
t
r
a
s
t
(
D
A
C
)
3
3
0
.
0
3
1
9
1
R
e
q
u
i
r
e
s
p
a
r
a
m
e
t
e
r
s
e
l
e
c
t
i
o
n
.
(
4
.
2
.
7
)
D
e
r
i
v
a
t
i
v
e
I
m
a
g
e
s
2
3
4
.
0
3
7
U
s
e
l
i
m
i
t
e
d
t
i
m
e
s
e
q
u
e
n
c
e
f
o
r
s
t
r
i
k
e
o
u
t
s
.
(
4
.
2
.
8
)
P
r
i
n
c
i
p
a
l
C
o
m
p
o
n
e
n
t
T
h
e
r
m
o
g
r
a
p
h
y
(
P
C
T
)
2
2
0
.
7
8
5
0
M
a
y
e
n
h
a
n
c
e
s
o
m
e
f
e
a
t
u
r
e
s
o
v
e
r
o
t
h
e
r
s
.
U
s
e
l
i
m
i
t
e
d
t
i
m
e
s
e
q
u
e
n
c
e
f
o
r
s
t
r
i
k
e
o
u
t
s
.
(
4
.
2
.
9
)
C
l
a
s
s
i
c
a
l
D
y
n
a
m
i
c
T
h
e
r
m
a
l
T
o
m
o
g
r
a
p
h
y
(
D
T
T
)
1
2
2
.
9
4
8
P
e
r
f
o
r
m
s
m
o
d
e
r
a
t
e
l
y
f
o
r
d
e
t
e
c
t
i
n
g
s
t
r
i
k
e
o
u
t
s
.
(
4
.
2
.
1
0
)
R
e
f
e
r
e
n
c
e
-
f
r
e
e
D
y
n
a
m
i
c
T
h
e
r
m
a
l
T
o
m
o
g
r
a
p
h
y
(
D
T
T
)
2
1
6
.
8
4
3
U
n
a
b
l
e
t
o
r
e
v
e
a
l
t
h
e
d
e
e
p
e
s
t
d
e
f
e
c
t
s
a
n
d
h
a
s
a
m
o
d
e
r
a
t
e
c
o
m
p
u
t
a
t
i
o
n
a
l
l
o
a
d
.
(
4
.
2
.
1
0
)
P
u
l
s
e
d
P
h
a
s
e
T
h
e
r
m
o
g
r
a
p
h
y
(
P
P
T
)
3
3
0
.
3
4
6
5
C
o
n
s
i
s
t
e
n
t
l
y
p
e
r
f
o
r
m
s
w
e
l
l
.
(
4
.
2
.
1
1
)
C
o
r
r
e
l
a
t
i
o
n
I
m
a
g
e
s
0
0
1
0
.
7
6
G
e
n
e
r
a
l
l
y
p
o
o
r
p
e
r
f
o
r
m
a
n
c
e
.
(
4
.
2
.
1
2
)
T
r
a
n
s
m
i
s
s
i
o
n
L
i
n
e
M
a
t
r
i
x
F
i
t
t
i
n
g
(
T
L
M
F
a
)
0
0
4
0
4
4
P
r
o
h
i
b
i
t
i
v
e
l
y
l
a
r
g
e
c
o
m
p
u
t
a
t
i
o
n
a
l
e
x
p
e
n
s
e
.
(
4
.
2
.
1
3
)
P
T
H
o
u
g
h
T
r
a
n
s
f
o
r
m
(
P
T
H
T
a
)
1
0
1
2
5
.
9
L
a
r
g
e
c
o
m
p
u
t
a
t
i
o
n
a
l
e
x
p
e
n
s
e
.
(
4
.
2
.
1
4
)
T
i
m
e
-
D
i

e
r
e
n
c
e
C
o
n
t
r
a
s
t
(
T
D
C
)
3
3
1
.
3
5
7
D
i

e
r
e
n
c
i
n
g
e
a
c
h
f
r
a
m
e
f
r
o
m
t
h
e

f
t
h
p
r
e
v
i
-
o
u
s
i
s
e

e
c
t
i
v
e
.
(
4
.
3
.
1
)
T
o
t
a
l
H
a
r
m
o
n
i
c
D
i
s
t
o
r
t
i
o
n
(
T
H
D
)
2
2
0
.
3
3
6
5
W
o
r
k
s
b
e
s
t
w
i
t
h
s
i
m
p
l
e
d
o
c
u
m
e
n
t
s
t
r
u
c
t
u
r
e
s
.
(
4
.
3
.
2
)
M
a
r
k
o
v
E
r
r
o
r
C
o
n
t
r
a
s
t
(
M
E
C
)
3
3
0
.
1
5
9
7
P
e
r
f
o
r
m
s
c
o
n
s
i
s
t
e
n
t
l
y
w
e
l
l
.
(
4
.
3
.
3
)
T
i
m
e
C
o
n
s
t
a
n
t
A
n
a
l
y
s
i
s
(
T
C
A
)
1
1
9
9
.
9
6
L
i
m
i
t
e
d
t
o
s
i
m
p
l
e
d
o
c
u
m
e
n
t
s
a
n
d
h
a
s
a
l
a
r
g
e
c
o
m
p
u
t
a
t
i
o
n
a
l
e
x
p
e
n
s
e
.
(
4
.
3
.
4
)
M
a
t
c
h
e
d
F
i
l
t
e
r
s
(
S
A
M
,
A
C
E
,
t
/
F
-
s
t
a
t
i
s
t
i
c
)
3
3
1
.
0
7
5
R
e
q
u
i
r
e
s
p
r
o
p
e
r
s
e
l
e
c
t
i
o
n
o
f
d
e
e
p
e
s
t
d
e
f
e
c
t
a
r
e
a
.
(
4
.
3
.
5
)
83
References
[1] F. C. Sham, N. Chen, and L. Long, Surface crack detection by ash thermography
on concrete surface, Insight, vol. 50, no. 5, pp. 240243, April 2008.
[2] J. Sham Fung Chu, Studies of Using Infrared Flash Thermography (FT) for Detec-
tion of Surface Cracks, Subsurface Defects and Water-Paths in Building Concrete
Structures. Ph.D. dissertation, City University of Hong Kong, July 2008.
[3] H. Nayeb-Hashemi, D. Swet, and A. Vaziri, New electrical potential method for
measuring crack growth in nonconducive materials, Measurements, vol. 36, pp. 121
129, 2004.
[4] M. A. Omar, Y. Zhou, R. Parvataneni, and E. Planting, Calibrated pulse-
thermography procedure for inspecting HDPE, Research Letters in Materials Sci-
ence, vol. 2008, no. 186427, p. 4, 2008.
[5] K. T. Wan and C. K. Y. Leung, Fiber optics sensor for the monitoring of mixed
mode cracks in structures, Sensors and Actuators, vol. 135, pp. 370380, 2007.
[6] C. Ibarra-Castanedo, M. Genest, P. Servais, X. P. V. Maldague, and A. Bendada,
Qualitative and quantitative assessment of aerospace structures by pulsed thermog-
raphy, Nondestructive Testing and Evaluation, vol. 22, no. 2-3, pp. 199215, June-
September 2007.
[7] P. Meinlschmidt, Thermographic detection of defects in wood and wood-based mate-
rials, 14th International Symposium of Nondestructive Testing of Wood, May 2005.
[8] M. Y. Y. Hung, Y. S. Chen, S. P. Ng, S. M. Shepard, Y. Hou, and J. R. Lhota,
Review and comparison of shearography and pulsed thermography for adhesive
bond evaluation, Optical Engineering, vol. 46, no. 5, p. 051007, 2007 [Online].
Available: http://link.aip.org/link/?JOE/46/051007/1.
[9] J. A. Schroeder, T. Ahmed, B. Chaudhry, and S. Shepard, Non-destructive
testing of structural composites and adhesively bonded composite joints: pulsed
thermography, Composites Part A: Applied Science and Manufacturing, vol. 33,
no. 11, pp. 15111517, 2002 [Online]. Available: http://www.sciencedirect.com/
science/article/B6TWN-47YXF1T-6/2/bc71b40b35ddc50e5ea83126db42480d.
[10] D. A. Scott, M. Dennis, N. Khandekar, J. Keeney, D. Carson, and L. S. Dodd, An
Egyptian cartonnage of the Graeco-Roman period: examination and discoveries,
Studies in Conservation, vol. 48, no. 1, pp. 4156, 2003.
[11] D. Bomford, Art in the Making: Underdrawings in Renaissance Paintings, ser. Na-
tional Gallery London Publications. New Haven: Yale University Press, 2002.
84
[12] B. Berrie, E. R. de la Rie, R. Homan, J. Tomlinson, T. Wiesel, and J. Winter,
Scientic Examination of Art: Modern Techniques in Conservation and Analysis.
Washingtion DC: National Academy of Sciences, 2002.
[13] L. MacDonald, Digital Heritage: Applying Digital Imaging to Cultural Heritage. Am-
sterdam: Elsevier, 2006.
[14] C. M. Falco, High resolution digital camera for infrared reectography, Review of
Scientic Instruments, vol. 80, no. 7, pp. 071 301071 309, July 2009.
[15] A. H. Bendada, S. Sfarra, D. Ambrosini, D. Paoletti, C. Ibarra-Castanedo, and
X. Maldague, Active thermography data processing for the NDT&E of frescoes, in
Quantitative Infrared Thermography 2010, pp. 961966, Quebec City, Canada, July
2010.
[16] G. M. Carlomagno and C. Meola, Comparison between thermographic techniques
for frescoes NDT, NDT & E International, vol. 35, no. 8, pp. 559565, 2002.
[17] K. Blessley, C. Young, J. Nunn, J. Coddington, and S. M. Shepard, The feasibility
of ash thermography for the examination and conservation of works of art, Studies
in Conservation, vol. 55, pp. 107120, 2010.
[18] D. Gavrilov, C. Ibarra-Castanedo, E. Maeva, O. Grube, X. Maldague, and R. Maev,
Infrared methods in noninvasive inspection of artwork, 9th International Conference
on Non-destructive Testing of Art, 2008.
[19] C. Ibarra-Castanedo, S. Sfarra, D. Ambrosini, D. Paoletti, A. Bendada, and
X. Maldague, Subsurface defect characterization in artworks by quantitative PPT,
Quantitative Infrared Thermography, 2008.
[20] D. Ambrosini, C. Daara, R. D. Biase, D. Paoletti, L. Pezzati, R. Bellucci, and F. Bet-
tini, Integrated reectography and thermography for wooden paintings diagnostics,
Journal of Cultural Heritage, vol. 11, no. 2, pp. 196204, 2010.
[21] M. Lettner and R. Sablatnig, Multispectral imaging for analyzing ancient
manuscripts, 17th European Signal Processing Conference, August 2009.
[22] K. Rapantzikos and C. Balas, Hyperspectral imaging: potential in non-destructive
analysis of palimpsests, in IEEE International Conference on Image Processing,
vol. 2, pp. II61821, Sept. 2005.
[23] G. A. Ware, D. M. Chabries, and R. W. Christiansen, Multispectral document en-
hancement: Ancient carbonized scrolls, IEEE Proceedings on Geoscience and Remote
Sensing Symposium, pp. 24862488, 2000.
[24] R. J. Easton, K. Knox, and W. Christens-Barry, Multispectral imaging of the
Archimedes palimpsest, in Applied Imagery Pattern Recognition Workshop, pp. 111
116, Oct. 2003.
[25] D. Obbink, A new Archilochus poem, Zeitschrift f ur Papyrologie und Epigraphik,
vol. 156, pp. 19, 2006.
85
[26] B. Zuckerman, Bringing the Dead Sea scrolls back to life: A new evaluation of
photographic and electric imaging of the Dead Sea scrolls, Dead Sea Discoveries,
vol. 3, no. 2, pp. 178207, 1996.
[27] G. Bearman, B. Zuckerman, K. Zuckerman, and J. Chiu, Multi-spectral digital imag-
ing of Dead Sea scrolls and other ancient documents, Jet Propulsion Laboratory,
Tech. Rep., 1993.
[28] J. Monaghan, G. Ware, J. Pohl, and S. Houston, A codex imaging project at the
Bodleian library: the recovery of lost mixtec writing, Foundation for the Advance-
ment of Meso-American Studies, 2004.
[29] U. Bergmann and K. T. Knox, Pseudo-color enhanced X-ray uorescence imaging of
the Archimedes Palimpsest, Proceedings Document Recognition and Retrieval, vol.
XVI, pp. 110, Jan. 2009.
[30] X. P. Maldague, Theory and Practice of Infrared Technology for Nondestructive Test-
ing. Hoboken: John Wiley & Sons, Inc., 2001.
[31] R. C. Gonzalez and R. E. Woods, Digital Image Processing. Upper Saddle River:
Pearson Education, Inc., 2008.
[32] J. Zalameda, N. Rajic, and W. Winfree, A comparison of image processing algorithms
for thermal nondestructive evaluation, SPIE Proc. Thermosense XXV, vol. 5073, pp.
374385, 2003.
[33] C. Ibarra-Castanedo, D. Gonzalez, M. Klein, M. Pilla, S. Vallerand, and X. Maldague,
Infrared image processing and data analysis, Infrared Physics & Technology, vol. 46,
pp. 7583, Apr. 2004.
[34] J. Ramirez-Granados, G. Paez, and M. Strojnik, Reconstruction and analysis of
pulsed thermographic sequences for nondestructive testing of layered materials, Ap-
plied Optics, vol. 49, no. 9, Mar. 2010.
[35] S. Shepard, J. Lhota, B. Rubadeux, T. Ahmed, and D. Wang, Enhancement and
reconstruction of thermographic data, Proceedings SPIE, vol. 4710, pp. 531535,
2002.
[36] S. M. Shepard, Temporal noise reduction, compression and analysis of
thermographic image data sequences, Patent, 02 2003, uS 6516084 [Online].
Available: http://www.patentlens.net/patentlens/patent/US_6516084/en/.
[37] M. Pilla, M. Klein, X. Maldague, and A. Salerno, New absolute contrast for pulsed
thermography, Proceedings of Quantitative Flash Thermography, pp. 5358, 2002.
[38] D. Gonzalez, C. Ibarra-Castanedo, M. Pilla, M. Klein, and X. Maldague, Automated
dierential absolute contrast, Quantitative Infrared Thermography, 2004.
[39] M. Susa, H. Benitez, C. Ibarra-Castanedo, H. Loaiza, H. Bendada, and X. Maldague,
Phase contrast using a dierentiated absolute contrast method, Quantitative In-
frared Thermography, vol. 3, pp. 219230, 2006.
86
[40] D. Gonzalez, C. Ibarra-Castanedo, M. Pilla, M. Klein, J. Lopez-Higuera, and
X. Maldague, Automatic interpolated dierentiated absolute contrast algorithm for
the analysis of pulsed thermographic sequences, in Quantitative Infrared Thermogra-
phy 7, D. Balageas, J. Buchlin, G. Carlomagno, and G. Busse, Eds., pp. H.16.1H.16.6,
Bruxelles, Belgique, July 2004.
[41] H. Benitez, C. Ibarra-Castanedo, A. Bendada, X. Maldague, H. Loaiza, and
E. Caicedo, Denition of a new thermal contrast and pulse correction for defect
quantication in pulsed thermography, Infrared Physics & Technology, vol. 51, pp.
160167, 2008.
[42] C. Ibarra-Castanedo, H. Benitez, X. Maldague, and A. Bendada, Review of thermal-
contrast-based signal processing techniques for the nondestructive testing and eval-
uation of materials by infrared thermography, International Workshop on Imaging
Nondestructive Evaluation, 07.
[43] H. Benitez, X. Maldague, C. Ibarra-Castanedo, H. Loaiza, A. Bendada, and
E. Caicedo, Modied dierential absolute contrast using thermal quadrupoles for
the nondestructive testing of nite thickness specimens by infrared thermography, in
Electrical and Computer Engineering, 2006. CCECE 06. Canadian Conference on,
pp. 10391042, may 2006.
[44] X. Maldague and S. Marinetti, Pulse phase infrared thermography, Journal of Ap-
plied Physics, vol. 79, no. 5, pp. 26942698, mar 1996.
[45] M. A. Omar and Y. Zhou, A quantitative review of three ash thermography
processing routines, Infrared Physics & Technology, vol. 51, no. 4, pp. 300
306, 2008 [Online]. Available: http://www.sciencedirect.com/science/article/
B6TJ9-4R0643G-2/2/1012d8792b0b432e3dfd2514330f125c.
[46] G. Busse, D. Wu, and W. Karpen, Thermal wave imaging with phase sensitive
modulated thermography, Journal of Applied Physics, vol. 71, no. 8, pp. 39623965,
apr 1992.
[47] X. Maldague, F. Galmiche, and A. Ziadi, Advances in pulsed phase thermography,
Infrared Physics & Technology, vol. 43, pp. 175181, 2002.
[48] C. Ibarra-Castanedo and X. Maldague, Defect depth retrieval from pulsed phase
thermographic data on plexiglas and aluminum samples, SPIE Proc. Thermosense
XXVI, vol. 5405, 2004.
[49] D. Gonzalez, C. Ibarra-Castanedo, F. Madruga, and X. Maldague, Dierentiated
absolute phase contrast algorithm for the analysis of pulsed thermographic sequences,
Infrared Physics & Technology, vol. 48, pp. 1621, 2006.
[50] M. Klein, A. Bendada, C. Ibarra-Castanedo, and X. Maldague, A hybrid pulsed
thermography processing technique for the depth estimation of subsurface defects
combining TSR and PPT, 10th International Conference on Quantitative Infrared
Thermography, 2010.
87
[51] S. M. Shepard. (2007, October) Flash thermography of aerospace composites
[Online]. Available: http://www.ndt.net/article/panndt2007/papers/132.pdf.
[52] S. M. Shepard, J. R. Lhota, B. A. Rubadeux, D. Wang, and T. Ahmed, Reconstruc-
tion and enhancement of active thermographic image sequences, Optical Engineering,
vol. 42, no. 5, May 2003.
[53] J. G. Sun, Analysis of pulsed thermography methods for defect depth prediction,
Journal of Heat Transfer, vol. 128, no. 4, pp. 329338, 2006.
[54] S. M. Shepard, Understanding ash thermography, Materials Evaluation, pp. 460
464, May 2006.
[55] S. M. Shepard, J. Hou, J. R. Lhota, and J. M. Golden, Automated processing of
thermographic deriviatives for quality assurance, Optical Engineering, vol. 46, no. 5,
May 2007.
[56] N. Rajic, Principal component thermography for aw contrast enhancement and
aw depth characterisation in composite structures, Composite Structures, vol. 58,
no. 4, pp. 521528, 2002 [Online]. Available: http://www.sciencedirect.com/
science/article/B6TWP-472841H-9/2/4c29e1491d0beb293d9a3811ed9498b8.
[57] N. Rajic, Principal component thermography, Defense Science and Technology
Organisation Victoria (Australia) Aeronautical and Maritime Research Laboratory,
Tech. Rep. 012-294, Apr. 2002.
[58] V. Vavilov, D. Nesteruk, V. Shiryaev, A. Ivanov, and W. Swiderski, Thermal (in-
frared) tomography: terminology, principal procedures, and application to nonde-
structive testing of composite materials, Russian Journal of Nondestructive Testing,
vol. 46, no. 3, pp. 151161, 2010.
[59] V. P. Vavilov, Dynamic thermal tomography: perspective eld of thermal NDT,
S. A. Semanovich, Ed., vol. 1313, no. 1, pp. 178182. SPIE, 1990 [Online]. Available:
http://link.aip.org/link/?PSI/1313/178/1.
[60] W. Swiderski, The characterization of defects in multi-layered composite materials
by thermal tomography methods, in Proceedings of the Tenth Annual Conference of
the Materials Research Society of Serbia, vol. 115, no. 4, pp. 800804, 2009.
[61] B. Sun, Q. Ma, and H. Zhao, Fitting-correlation analysis of pulsed thermographic
sequence data, in Proceedings of the 2007 IEEE International Conference on Mecha-
tronics and Automation, Aug. 2007.
[62] M. Klein, C. Ibarra-Castanedo, A. H. Bendada, and X. Maldague, Thermographic
signal processing through correlation operators in pulsed thermography, in Ther-
mosense XXX, SPIE Defense and Security Symposium, V. P. Vavilov and D. D.
Burleigh, Eds., vol. 6939, p. 693915, Orlando, Florida, USA, Vol. 6939, March 16-20
2008.
88
[63] D. Gonzalez, C. Ibarra-Castanedo, J. Lopez-Higuera, and X. Maldague, New algo-
rithm based on the Hough transform for the anlaysis of pulsed thermographic se-
quences, Nondestructive Testing and Evaluation International, vol. 39, pp. 617621,
2006.
[64] C. Ibarra-Castanedo, D. Gonzalez, F. Galmiche, X. Maldague, and A. H. Bendada,
Discrete signal transforms as a tool for processing and analyzing pulsed thermo-
graphic data, in SPIE - The International Society for Optical Engineering, Ther-
mosense XXVIII, vol. 6205, pp. 620 5141620 51412, Orlando, FL, 2006.
[65] D. de Cogan, A. Soulos, and K. Chichlowski, Sub-surface feature location and identi-
cation using inverse TLM techniques, Microelectronics Journal, vol. 29, pp. 21522,
1998.
[66] D. de Cogan, W. J. OConnor, and S. Pulko, Transmission Line Matrix in Computa-
tional Mechanics. Boca Raton: Taylor and Francis, 2006.
[67] D. Crowther, L. Favro, P. Kuo, and R. Thomas, Inverse scattering algorithm applied
to infrared thermal waves, J. Appl. Phys., vol. 74, no. 9, pp. 58285835, Nov 1993.
[68] Favro, L. D., Han, X., Kuo, P. K., and Thomas, R. L., Improving the resolution of
pulsed thermal wave images with a simple inverse scattering technique, Journal De
Physique IV France, vol. 04, pp. C7545C7550, 1994.
[69] L. Favro, D. Crowther, P. Kuo, and R. Thomas, Inversion of pulsed thermal-wave
images for defect sizing and shape recovery, Thermosense XIV: An International
Conference on Thermal Sensing and Imaging Diagnostic Applications, vol. 1682, pp.
178181, 1992.
[70] N. S. Goel and F. Gang, A simple method for pulse-echo thermal wave imaging
of arbitrary shaped subsurface scatterers in heterogeneous materials, International
Communications in Heat Mass Transfer, vol. 23, no. 1, pp. 4554, 1996.
[71] R. Thomas, L. Favro, and P. Kuo, Thermal wave imaging of hidden corrosion,
Wayne State University, Tech. Rep. F49620-96-1-0166, Mar. 1998.
[72] M. Omar, H. Hassan, and K. Saito, Optimizing thermography depth probing with a
dynamic thermal point spread function, Infrared Physics & Technology, vol. 46, pp.
506514, 2005.
[73] T. DOrazio, C. Guaragnella, M. Leo, and P. Spagnolo, Defect detection in aircraft
composites by using a nerual appoach in the analysis of thermographic sequences,
Nondestructive Testing and Evaluation International, vol. 38, pp. 665673, 2005.
[74] D. Maillet, J. Batsale, A. Bendada, and A. Degiovanni, Methodes integrales et con-
trole non destructif par thermographie infrarouge, Revue Generale de Thermique,
vol. 35, pp. 1427, 1996.
[75] A. Bendada, F. Erchiqui, and M. Lamontagne, Pulsed thermography in the evalu-
ation of an aircraft composite using 3D thermal quadrupoles and mathematical per-
turbations, Inverse Problems, vol. 21, pp. 857577, 2005.
89
[76] V. Vavilov, IR Thermography and Nondestructive Testing. Moscow: Spektr, 2009.
[77] J. Sun, Method for determining defect depth using thermal imaging, Patent, 04
2002, US 6542849.
[78] N. Rajic, A quantitative approach to active thermographic inspection for material
loss evaluation in metallic structures, Research in Nondestructive Evaluation, vol. 12,
no. 2, pp. 119131, 2000.
[79] V. Vavilov, X. Maldague, B. Dufort, and A. Ivanov, Adaptive thermal tomography
algorithm, Proceedings SPIE Thermosense-XV, no. 1933, pp. 166173, 1993.
[80] H. Benitez, H. Loaiza, E. Caicedo, C. Ibarra-Castanedo, A. Bendada, and
X. Maldague, Defect characterization in infrared non-destructive testing with learn-
ing machines, Nondestructive Testing & Evaluation International, vol. 42, pp. 630
643, 2009.
[81] I. Abdel-Qader, S. Yohali, O. Abudayyeh, and S. Yehia, Segmentation of thermal im-
ages for non-destructive evaluation of bridge decks, Nondestructive Testing & Eval-
uation International, vol. 41, no. 5, pp. 395405, 2008.
[82] V. F. Bakirov, R. A. Kline, and W. P. Winfree, Multiparameter thermal tomogra-
phy, AIP Conference Proceedings, vol. 700, no. 1, pp. 461468, 2004.
[83] V. Bakirov, R. Kline, and W. Winfree, Diusion based thermal tomography, Non-
destructive Evaluation and Reliability of Micro and Nonomaterial Systems, vol. 4703,
pp. 219225, 2002.
[84] V. F. Bakirov and R. A. Kline, Diusion-based thermal tomography, Journal of
Heat Transfer, vol. 127, no. 11, pp. 12761279, 2005.
[85] F. Ammirato and P. Zayicek, Infrared thermography eld application guide, Electric
Power Research Institute, Inc., Palo Alto, CA, Tech. Rep. TR-107142, Jan. 1999.
[86] M. Klein, C. Ibarra-Castanedo, X. Maldague, and A. H. Bendada, IR-View: a
straightforward graphical user interface for basic and advanced signal processing of
thermographic infrared sequences, in Thermosense XXX, SPIE Defense and Security
Symposium, V. P. Vavilov and D. D. Burleigh, Eds., vol. 6939, p. 693914, Orlando,
Florida, USA, Vol. 6939, March 16-20 2008.
[87] S. Shepard, J. Lhota, and T. Ahmed, Flash thermography contrast model based on
IR camera noise characteristics, Nondestructive Testing and Evaluation, vol. 22, no.
2-3, pp. 113126, 2007.
[88] V. Vavilov, Pulsed thermal NDT of materials: back to the basics, Nondestructive
Testing and Evaluation, vol. 22, no. 2-3, pp. 177197, June 2007.
[89] P. Zayicek, Infrared thermography guide (revision 3), Electric Power Research In-
stitute, Inc., Tech. Rep. 1006534, 2002.
90
[90] X. Maldague, Nondestructive Evaluation of Materials by Infrared Thermography.
London: Springer-Verlag, 1993.
[91] J. Krapez, D. Balegeas, A. Deom, and F. Lepoutre, Early detection by stimulated in-
frared thermography, in Advances in Signal Processing for Nondestructive Evaluation
of Materials, X. Maldague, Ed., pp. 303321. The Netherlands: Kluwer Academic
Published, 1994.
[92] J. R. Lhota and S. M. Shepard, Critical timing issues in ash thermography, D. O.
Thompson and D. E. Chimenti, Eds., vol. 760, no. 1, pp. 16541660. AIP, 2005
[Online]. Available: http://link.aip.org/link/?APC/760/1654/1.
[93] U. Netzelmann, Optimized temporal excitation proles in active thermography,
Analytical Sciences, vol. 17, pp. 432435, Apr. 2001.
[94] X. Maldague, Z. Ziadi, and M. Klein, Double pulse infrared thermography, Nonde-
structive Testing and Eval. Int., vol. 37, pp. 559564, 2004.
[95] A. Daniels, Optimization of pulsed temporal stimulation for thermal nondestructive
testing, in Thermosense XX SPIE, J. Snell and R. Wurzbach, Eds., vol. 3361, pp.
254265, 2001.
[96] H. Carslaw and J. Jaeger, Conduction of Heat in Solids, 2nd ed. Oxford: Claredon
Press, 1986.
[97] K. Wan and Q. Ma, Data-tting reconstruction for defect inspection of airplane
aluminum structure in infrared thermographic NDT, in Industrial Electronics and
Applications, 2009. ICIEA 2009. 4th IEEE Conference on, pp. 691696, may 2009.
[98] H. Ringermacher, D. Mayton, D. Howard, and B. Cassenti, Towards a at-bottom
hole standard for thermal imaging, Review of Progress in Quantitative Nondestruc-
tive Evaluation, vol. 17, pp. 4259, 1998.
[99] M. Omar, M. Hassan, K. Saito, and R. Alloo, IR self-referencing thermography for
detection of in-depth defects, Infrared Physics & Technology, vol. 46, pp. 283289,
2005.
[100] T. A. Griths, Enhancing multispectral imagery of ancient documents, Masters
thesis, Utah State University, 2011.
[101] B. Foy, Overview of target detection algorithms for hyperspectral data, Los Alamos
National Laboratory, Tech. Rep. LA-UR-09-00593, June 2009.
[102] D. Manolakis, Taxonomy of detection algorithms for hyperspectral imaging applica-
tions, Optical Engineering, vol. 44, no. 6, pp. 066 4031066 40311, June 2005.
[103] M. E. Smith, Codex selden: A manuscript from the Valley of Nochixlan, in The
Cloud People: Divergent evolution of the Zapotec and Mixtec civilizations, K. Flannery
and J. Marcus, Eds., pp. 248255. New York: Academic Press, 1984.
91
[104] M. E. Smith, Why the second Codex Selden was painted, in Caciques and their
people, J. Marcus and J. F. Zeitlin, Eds., vol. 89, pp. 111142. Anthropological
Papers, 1994.
[105] P. Dark and J. Plesters, The palimpsests of Codex Selden: Recent attempts to reveal
the covered pictographs, 33rd Annual International Congress of Americanists, vol. 2,
pp. 530539, 1958.
[106] G. A. Ware, Codex selden feasibility study: multispectral and infrared images,
Brigham Young University, Tech. Rep., 2005.
92
Appendices
93
Appendix A
Pulsed Thermography Toolbox
A.1 Pulsed Thermography Toolbox (PTT) Introduction
The Pulsed Thermography Toolbox (PTT) was developed as part of this research to aid
in the application of pulsed thermography to ancient documents and archaeological artifacts.
The toolbox provides the processing techniques discussed in Chapter 4 in a easy-to-use
graphical user interface (GUI). The code was developed and implemented in MATLAB
R2010b. The main GUI interface can be seen in Figure A.1. The GUI subcomponents will
be discussed in detail throughout this chapter.
A.2 Load Images
The load image sub-panel is shown in Figure A.2. Three data types can be selected:
tagged image le format (.tif or .ti ), a proprietary binary format develop by the Space
Dynamics Lab (.sbf ), or a MATLAB binary (.mat) le type.
After selecting the desired image format, click the Load Images button and a le
selection window (Figure A.3) will appear.
If the TIF radial button is selected, it will prompt selection of a directory containing
the .tif les and will load the les within that directory in alphabetical order. If the
SDL Binary or MATLAB Binary (.mat) options are selected, the le selection window will
prompt selection of select a single .sbf or .mat le, respectively. Note that in order for the
MATLAB binary to be properly loaded, it must contain two matrices. The rst matrix
must be called imageCube and contain the pulsed thermography data in a MNT array,
where MN are the dimensions of each frame and T is the number of frames. In addition,
the MATLAB binary must contain a single MN matrix containing the pre-ash image
94
Fig. A.1: Pulsed Thermography Toolbox main GUI.
Fig. A.2: Load images panel.
Fig. A.3: Select les interface.
95
with the variable name preash. After the data are loaded, the rst image of the data set
will be shown on the Display Panel (Figure A.4).
The Display Panel features a slider that controls which frame is currently displayed.
The Play All button iterates through every frame showing the video sequence. Finally,
the image number is displayed in the upper left, and the image tag information is displayed
in the upper right.
A.3 Manually Manipulate Data Panel
The Manually Manipulate Data Panel is shown in Figure A.5 and provides a variety
of features for exploring and manipulating the data manually. The Explore Data button
brings up the Explore Data Sub-GUI as discussed in Section A.8.
The Crop Images button allows each frame in the data set to be cropped. After clicking
the Crop Images button, the image corresponding to the number in the FlashPoint edit
box will be displayed allowing selection of a rectangular region on the image displayed in
the Display Panel. After selecting the region to keep, right-click on the image and select
the Crop Image item from the drop-down menu. Each image frame will be cropped to the
specied dimensions.
The Select Colormap drop-down box allows the user to select the colormap to be
used in creating the false color images (see Section 4.2.1). To view the original grayscale
intensity images, select the Grayscale option from the drop-down box.
The FlashPoint and RemoveEnd edit boxes are initialized with the length of the data.
These boxes allow for the data set to be reduced in length. Changing the number in the
FlashPoint edit box allows removal of any images before that point (pre-ash images).
By changing the number in the RemoveEnd edit box, the total number of frames can be
reduced. After entering the desired parameters, press the Remove Images button to apply
the actions. Note that when changing the Flashpoint, the third previous image will be
saved as the pre-ash image used for future calculations. Finally, the Est. FlashPoint
button can be used to automatically detect the ash point, give an estimation of the length
96
Fig. A.4: Display panel.
Fig. A.5: Manually Manipulate Data Panel.
97
of data needed, and update these values in the FlashPoint and RemoveEnd edit boxes,
respectively.
A.4 Pre-Processing Panel
The Pre-Processing Panel, shown in Figure A.6, provides useful pre-processing routines.
Recommended settings are pre-selected. The Remove Surface Effusivities subtracts
the pre-ash image from each image in the data set to reduce the eects of dierences in
surface eusivity values. The Spatial Noise Reduction selects the image lters discussed
in Section 4.2.4. Both the lters use a 33 window. The variance of the Gaussian lter,
which aects the amount of blurring, can be changed using the slider box labeled Variance.
The Normalize Cube checkbox allows for the entire data set to be re-normalized between 0
and 1. This is useful when the data set does not ll the entire range due to the median lter
removing dead pixels or other abnormalities. Finally, the Signal Reconstruction provides
a choice between the synthetic signal reconstruction techniques discussed in Section 4.2.5.
When Layered Reconstruction is selected, the number of layers to use in reconstruction
can be changed using the Num. Layers slider. In addition, the reconstructed signal can
be modied using the Interpolation edit box. An interpolation factor of 2 will up-sample
the signal by double. Similarly, an interpolation factor of 0.5 will down-sample the signal
by one half. When all the chosen pre-processing techniques have been selected, click the
Pre-Process button to apply them to the data set.
Fig. A.6: Pre-Processing Panel.
98
A.5 Processing Panel
The Processing Panel is shown in Figure A.7 and allows for the operator to choose
a desired processing algorithm as discussed in Chapter 4. When the desired algorithm is
chosen, click the Process button to apply the algorithm to the data set. One of the options
in the drop-down box is ### Sandbox ###. When this option is selected, the Sandbox
drop-down menu is enabled. This drop-down menu is automatically populated with any
functions located in the sandbox/ directory. Choose any of the functions in the sandbox
drop-down menu and click Process to run the selected algorithm. Note that any algorithm
placed in the sandbox/ directory is required to have a standard header with a single input
and two outputs. The input is an image cube of the input data with dimensions MNT.
The outputs are a processed data cube MNR and a cell array of string image descriptors
of length R. This facilitates straightforward testing of new processing routines which can
be easily incorporated into the existing GUI.
A.6 Post-Processing Panel
The Post-Processing Panel is seen in Figure A.8 and provides the three contrast stretch-
ing routines discussed in Sections 4.2.2 and 4.2.3. After selecting the desired post-processing
technique, click Post-Process button to apply to the data set.
A.7 Other GUI Sub-Components
There are several other GUI sub-components. The Which Cube panel allows for the
selection between data sets at dierent stages of processing. The Enable Multicore
Processing makes it possible for MATLAB to use four processors when processing the
Fig. A.7: Processing Panel.
99
Fig. A.8: Post-Processing Panel.
data. The Adjust Contrast button displays the histogram of the currently displayed im-
age frame and allows for manual histogram manipulation. The Open IR View passes the
data to the IR View toolbox, a previously existing open-source toolbox for processing ther-
mal data. Note that the IR View toolbox was made for a previous version of MATLAB
and not all of the functions work correctly. The Write To File outputs the currently se-
lected data to le in the results/[algorithm name]/ directory in .tif format. The lenames
are generated from the image tag cell array. Finally, the Exit button closes the GUI and
disables the multi-core processing nodes created by MATLAB.
A.8 Explore Data Sub-GUI
The Explore Data Sub-GUI is launched when the Explore Data button is selected in
the Manually Manipulate Data Panel. This sub-GUI is shown in Figure A.9 and allows for
the visualization of the temperature decay sequences. It contains four image panels: the
current frame (upper right), the selected pixel time sequence (upper left), the selected pixel
time sequence in log domain (lower left), and the Fourier Transform of the selected pixel time
sequence (lower right). Four models for reconstructing or tting to the data are available for
analysis: thermal signal reconstruction, Markov model, layered reconstruction, and reverse
transmission line matrix models. Also, the optional Show Neighbors check box replaces
the Fourier transform with the selected pixels time sequence and the time sequence of its
immediate spatial neighbors. Selecting the Run button will allow the operator to choose a
desired pixel to plot. The Save Series button will save the currently displayed pixel time
series to a MATLAB .mat binary. Finally, the Exit button will close the sub-GUI.
100
Fig. A.9: Explore Data Sub-GUI.
101
Appendix B
Code Listings
The listings for the main processing techniques are given.
B.1 Contrast Denitions & Dierential Absolute Contrast (DAC)
function [ outCube outTags CtrstDef ] = ContrastImages ( imageCube
, T to )
% CONTRASTIMAGES Al l ows s e l e c t i o n bet ween Absol ut e , Running ,
Normal ized ,
% and Standard c ont r as t d e f i n i t i o n s . In addi t i on , d i f f e r e n t i a l
ab s ol ut e
% c ont r as t (DAC) i s al s o s e l e c t a b l e .
%
% [U V W] = Cont rast Images (X, Y) Input X i s a 3D
% imageCube of di mensi on MxNxT, where MxN are t he frame s i z e and
T i s t he
% number of f rames . Input Y i s a pref l a s h frame of di mensi ons
MxN. Output
% U i s t he proces s ed image cube , V i s t he image t ag i nf ormat i on
as s oc i at e d
% wi t h each frame , and W i s t he s e l e c t i o n of whi ch c ont r as t
d e f i ni t i o n was
% chosen .
%
% Ref erence Paper ( s ) :
102
% 1) X. P. Maldague , Theory and Pr act i ce of I nf r ar ed Technol ogy
f or
% Nondes t r uct i ve Test i ng . John Wil ey & Sons , Inc . , 2001.
% 2) M. Pi l l a , M. Kl ei n , X. Maldague , and A. Sal erno , New
ab s ol ut e
% c ont r as t f or pul s ed thermography , Proceedi ngs of Quant i t at i ve
Fl ash
% Thermography , pp . 5358, 2002.
%
%
% See al s o GETPOSITION.
%
% PULSED THERMOGRAPHY PROCESSING TOOLBOX
% Author : Cory Andrew Larsen
f pri ntf ( Running Contrast Images Al gori thm . . . \ n )
[ Nx Ny Nt ] = si ze ( imageCube ) ;
f pri ntf ( Se l e c t De f i ni t i on of Contrast to Use : \ n ) ;
f pri ntf ( 1) Absol ute Contrast \n ) ;
f pri ntf ( 2) Running Contrast \n ) ;
f pri ntf ( 3) Normal i zed Contrast ( us i ng f i n a l val ue ) \n ) ;
f pri ntf ( 4) Normal i zed Contrast ( us i ng maximal val ue ) \n ) ;
f pri ntf ( 5) Standard Contrast \n ) ;
f pri ntf ( 6) Di f f e r e nt i a t e d Absol ute Contrast (DAC) \n ) ;
f pri ntf ( \n ) ;
s e l e c t = input ( Which De f i ni t i on to Use?\n ) ;
103
i f ( s e l e c t < 1 | | s e l e c t > 6)
f pri ntf ( Error : I nval i d Se l e c t i o n . Choose a number between 1
& 7\n ) ;
s e l e c t = input ( Which De f i ni t i on to Use?\n ) ;
end
i f ( s e l e c t < 6)
f pri ntf ( Choose Background Pi xe l s : ) ;
f pri ntf ( Se l e c t 5 Poi nts . . . ) ;
Tsnd = 0;
Tsnd to = 0;
for nn=1:5
h = i mpoi nt ;
pos = ge t Pos i t i on ( h) ;
py = round( pos ( 1) ) ;
px = round( pos ( 2) ) ;
Tsnd = Tsnd + imageCube ( px , py , : ) ;
Tsnd to = Tsnd to + T to ( px , py) ;
end
Tsnd = permute ( Tsnd / 5 , [ 3 , 1 , 2 ] ) ;
Tsnd to = Tsnd to /5;
f pri ntf ( done . \ n ) ;
end
i f ( s e l e c t == 6)
tp = input ( Enter ti me i ndex of when f i r s t de f e c t appears :
) ;
del t a Tt p = imageCube ( : , : , tp ) T to ;
104
end
ti c ;
% Def i ne Time I ns t ances
T f = imageCube ( : , : , end) ;
T tm = imageCube ( : , : , 1 ) ;
% I n i t i a l i z e Arrays
AbsCtrst = zeros (Nx, Ny, Nt ) ; RunCtrst = AbsCtrst ; NormCtrstA =
AbsCtrst ;
NormCtrstB = AbsCtrst ; StdCtrst = AbsCtrst ; DAC = AbsCtrst ; CM =
AbsCtrst ;
for t t =1:Nt
T = imageCube ( : , : , t t ) ;
swi t ch s e l e c t
cas e 1
% Absol ut e Cont rast %
Ca = TTsnd( t t ) ;
i dx = Ca < 0;
Ca( i dx ) = 0;
AbsCtrst ( : , : , t t ) = Ca ;
cas e 2
% Running Cont rast %
Cr = (TTsnd( t t ) ) /Tsnd( t t ) ;
i dx = Cr < 0;
Cr ( i dx ) = 0;
RunCtrst ( : , : , t t ) = Cr ;
105
cas e 3
% Normal i zed Cont rast A%
Cn1 = T. / T f Tsnd( t t ) /Tsnd(end) ;
NormCtrstA ( : , : , t t ) = Cn1 ;
cas e 4
% Normal i zed Cont rast B%
Cn2 = T. / T tm Tsnd( t t ) /Tsnd( 1) ;
NormCtrstB ( : , : , t t ) = Cn2 ;
cas e 5
% Standard Cont rast %
Cs = (TT to ) / ( Tsnd( t t ) Tsnd to ) ;
StdCtrst ( : , : , t t ) = Cs ;
cas e 6
% Di f f e r e nt i a t e d Absol ut e Cont rast (DAC) %
del t a T = T T to ;
DAC( : , : , t t ) = del t a T sqrt ( tp/ t t ) del t a Tt p ;
end
end
% St ore Data f or Output %
swi t ch s e l e c t
cas e 1
outCube = AbsCtrst ;
CtrstDef = Abs ;
cas e 2
outCube = RunCtrst ;
CtrstDef = Run ;
cas e 3
106
outCube = NormCtrstA;
CtrstDef = Normf ;
cas e 4
outCube = NormCtrstB;
CtrstDef = Normi ;
cas e 5
outCube = StdCtrst ;
CtrstDef = Std ;
cas e 6
outCube = DAC;
CtrstDef = [ DAC i nt2str ( tp ) ] ;
end
outTags = genNumTags ( outCube ) ;
toc ;
Di spl ayTi me ( toc ) ;
f pri ntf ( Contrast Images Complete ! \ n ) ;
end
B.2 Thermal Signal Reconstruction (TSR) and Derivative Images
function [ f df ddf ] = t s r ( imageCube , or der )
% TSR Thermal Si gnal Reconst ruct i on al gor i t hm f or pul s ed
thermography
% dat a . The al gor i t hms removes t emporal noi s e i n t he dat a by
usi ng a l og
107
% t r ans f ormat i on and a pol ynomi al f i t . I t t hen t akes t he f i r s t
and second
% d e r i v a t i v e s of t he pol ynomi al f unc t i ons t o i ncr eas e t he
c ont r as t of
% d e f e c t s .
%
% [ F DF DDF] = TSR(X) Process 3D image cube X and r e s u l t s
% i n F, DF, and DDF. F, DF, and DDF are t he r econs t r uct ed images
, t he
% f i r s t d e r i v a t i v e images , and t he second d e r i v a t i v e images ,
r e s p e c t i v e l y .
% X, F, DF, and DDF are U by V by W, where U and V are t he
di mensi ons of
% each image frame and W i s t he number of f rames .
%
% [ F DF DDF] = TSR(X,N) Process X usi ng an Nth order pol ynomi al
f i t .
% Def aul t i s 6 t h order .
%
% Ref erence Paper ( s ) :
% 1) S. M. Shepard , Temporal noi s e reduct i on , compressi on and
anal y s i s
% of t hermographi c image dat a sequences , Patent , 02 2003 , US
6516084.
% 2) S. M. Shepard , J . R. Lhota , B. A. Rubadeux , D. Wang, and T.
Ahmed,
% Reconst ruct i on and enhancement of ac t i v e t hermographi c image
se
108
% quences , Opt i cal Engi neeri ng , v ol . 42 , no . 5 , May 2003.
%
% See al s o POLYFIT,POLYVAL.
%
% PULSED THERMOGRAPHY PROCESSING TOOLBOX
% Author : Cory Andrew Larsen
ti c ;
% Get Dimensions
[ Nx Ny Nt ] = si ze ( imageCube ) ;
M = NxNy;
imageMat = reshape( imageCube ,M, Nt ) ;
% nth order pol ynomi al f or c o e f f i c i e n t s
i f ( nargin>2)
nt h or der = or der ;
el se
nt h or der =6;
end
% de c l ar e ti me
t =(1: Nt ) ;
%% Take Nat ural Logs of p i x e l and ti me dat a
% I ( t ) i , j = l n [ I ( t ) i , j ]
l o g I = zeros (M, Nt ) ;
par f or nn=1:M
109
l o g I ( nn , : ) = log ( imageMat ( nn , : ) ) ;
end
l ogt=log ( t ) ;
f pri ntf ( Smoothing Data . . . \ n ) ;
%% Least Squares Fi t
% l n [ I ( t ) i , j ] = a0 + a1 l n ( t ) + a2 l n ( t ) 2 + . . . + an l n ( t ) n
I c o e f f = zeros (M, nt h or der +1) ;
par f or nn=1:M
I c o e f f ( nn , : ) = pol yf i t ( l ogt , l o g I ( nn , : ) , nt h or der ) ;
end
%% Regenerat e f ( t )
f = zeros (M, Nt ) ;
par f or nn=1:M
f ( nn , : ) = exp( polyval ( I c o e f f ( nn , : ) , l ogt ) ) ;
end
f pri ntf ( Take De r i vat i ve s . . . \ n ) ;
%% Find f ( t )
h = 1. / t ;
df = zeros (M, Nt ) ;
g = zeros (M, Nt ) ;
for nn=1:M
I c o e f f ( nn , : ) = fl i pud ( I c o e f f ( nn , : ) ) ;
for i =1: nt h or der
g ( nn , : ) = g ( nn , : ) + i I c o e f f ( nn , i +1) l ogt . ( i 1) ;
110
end
df ( nn , : ) = h. g ( nn , : ) . f ( nn , : ) ;
end
%% Find f ( t )
ddf = zeros (M, Nt ) ;
gp = zeros (M, Nt ) ;
for nn=1:M
for i =2: nt h or der
gp( nn , : ) = gp( nn , : ) + i ( i 1) I c o e f f ( nn , i +1) l ogt . ( i
2) ;
end
ddf ( nn , : ) = h . 2 . f ( nn , : ) . ( g ( nn , : ) .2g ( nn , : ) +gp( nn , : ) ) ;
end
% Transform back t o st andard form
f = imnorm( reshape( f , [ Nx Ny Nt ] ) ) ;
df = imnorm( reshape( df , [ Nx Ny Nt ] ) ) ;
ddf = imnorm( reshape( ddf , [ Nx Ny Nt ] ) ) ;
toc ;
Di spl ayTi me ( toc ) ;
f pri ntf ( TSR Complete ! \ n ) ;
111
B.3 Principal Component Thermography (PCT)
function [ outCube outTags ] = pct ( imageCube )
% PCT Pr i nc i pal Component Thermography al gor i t hm f or pul s ed
thermography
% dat a . The al gor i t hm uses s i ng ul ar val ue decomposi t i on t o reduce
t he dat a
% t o a s t a t i s t i c a l r e pr e s e nt at i on of s p a t i a l and t emporal
v ar i at i ons
% r e l a t i ng t o t he under l yi ng s t r uc t ur e of t he sampl e .
%
% [Y TAGS] = PCT(X) Process 3D image cube X and r e s u l t s i n Y, a
3D image
% cube cont ai ni ng t he f i r s t 10 empi r i cal or t hogonal f unc t i ons (
EOF) . X
% i s U by V by W, where U and V are t he di mensi ons of each image
frame and
% W i s t he number of f rames . Y i s U by V by 10. TAGS i s a vect or
of c e l l
% arrays cont ai ni ng a t e x t de s c r i pt i on of each out put frame .
%
% Ref erence Paper ( s ) :
% 1) N. Raj i c . Pr i nc i pal component thermography f or f l aw c ont r as t
enhancement
% and f l aw dept h c har ac t e r i s at i on i n composi t e s t r uc t ur e s .
Composite St r uct ur es ,
% 58(4) :521 528 , 2002.
%
112
% 2) N. Raj i c . Pr i nc i pal Component Thermography . Tech not e DSTO
TR1298
%
% See al s o SVD, FIXCUBE.
%
% PULSED THERMOGRAPHY PROCESSING TOOLBOX
% Author : Cory Andrew Larsen
ti c ;
f pri ntf ( Pr i nc i pal Component Thermography (PCT) Al gori thm
Begi nni ng . . . \ n ) ;
addpath ( . . / l i b / ) ;
% f i nd di mensi ons
[ Nx Ny Nt ] = si ze ( imageCube ) ;
M = NxNy;
N = Nt ;
% Reshape i nt o 2D MN array
A = reshape( imageCube ,M, N) ;
cl ear imageCube
% St andardi z e
mew = mean(A) ;
si gma = std(A) ;
for m=1:M
A( : ,m) = (A( : ,m) mew(m) ) / sigma (m) ;
end
cl ear mew sigma
113
% Make sure t her e are no NaN or INF i n Cube
A = f i xCube (A) ;
% Perform s i ng ul ar val ue decomposi t i on (SVD)
[ U Gamma V] = svd(A , econ ) ;
% Transform U i nt o o r i g i na l cube shape
outCube = reshape(U, Nx, Ny, Nt ) ;
outTags = genNumTags ( outCube ) ;
toc ;
Di spl ayTi me ( toc ) ;
f pri ntf ( Pr i nc i pal Component Thermography (PCT) Al gori thm
Complete ! \ n ) ;
B.4 Dynamic Thermal Tomography (DTT)
function [ outCube outTags ] = dt t ( imageCube )
% DTT Dynamic Thermal Tomography Al gori t hm . Performs bot h t he
c l a s s i c a l and t he
% r ef er ence f r e e al gor i t hm v ar i ant s .
%
% [Y T] = IMNORM(X) Processes i nput 3D image cube X, and r et ur ns
t he
% proces s ed 3D image cube Y and image de s c r i pt i on t ags i n T. X
and Y are
114
% U by V by W matri x , where U and V are t he di mensi ons of each
image frame
% and W i s t he number of f rames ( f or Y, W=4) . T i s a vect or of
c e l l arrays cont ai ni ng
% s t r i ng s de s c r i b i ng each image frame and i s of l e ng t h 4.
%
% Ref erence Papers :
% 1) V. Vavi l ov , D. Nesteruk , V. Shi ryaev , A. Ivanov , and W.
Swi derski ,
% Thermal ( i nf r ar e d ) tomography : Terminol ogy , p r i nc i p a l
procedures , and
% appl i c at i on t o nondes t r uc t i v e t e s t i ng of composi t e mat eri al s ,
Russi an
% Journal of Nondes t r uct i ve Test i ng , v ol . 46 , no . 3 , pp . 151161,
2010.
% 2) V. P. Vavi l ov , Dynamic t hermal tomography : pe r s pe c t i v e
f i e l d of
% t hermal ndt , S. A. Semanovich , Ed. , v ol . 1313 , no . 1. SPIE,
1990 , pp .
% 178182.
% 3) W. Swi derski , The c har ac t e r i z at i on of d e f e c t s i n mul t i
l ayer ed compos
% i t e mat er i al s by t hermal tomography methods , i n Proceedi ngs of
t he
% Tenth Annual Conf erence of t he Mat er i al s Research Soci et y of
Serbi a ,
% v ol . 115 , no . 4 , 2009.
%
115
%
% See al s o POLYFIT, IMPOINT, IMNORM
%
% PULSED THERMOGRAPHY PROCESSING TOOLBOX
% Author : Cory Andrew Larsen
addpath ( . . / l i b / ) ;
f pri ntf ( Running Dynamic Thermal Tomography (DTT) Al gori thm . . . \ n
) ;
% t i c ;
% f i nd di mensi ons
[ Nx Ny Nt ] = si ze ( imageCube ) ;
% Normal i ze wi t h r e s pe c t t o f i r s t frame
Tn = zeros (Nx, Ny, Nt ) ;
for nn=1:Nt
Tn( : , : , nn) = imageCube ( : , : , nn) . / imageCube ( : , : , 1 ) ;
end
cl ear imageCube
%% Cl a s s i c a l Approach
%Se l e c t r ef e r e nc e poi nt s and average
f pri ntf ( Se l e c t 5 Background Poi nts . . . ) ;
Tref = 0;
for nn=1:5
h = i mpoi nt ;
pos = ge t Pos i t i on ( h) ;
py = round( pos ( 1) ) ;
116
px = round( pos ( 2) ) ;
Tref = Tref + Tn( px , py , : ) ;
end
% Tref = permute ( Tref , [ 3 , 1 , 2 ] ) /5;
Tref = Tref /5;
f pri ntf ( done . Pr oces s i ng . . . \ n ) ;
ti c ;
% Cal cul at e d i f f e r e nc e bet ween dat a and r e f e r e nc e
% Find Maxigram and correspondi ng ti me gram
Cl assi cal Maxi gram = zeros (Nx, Ny) ;
Cl assi cal Ti megram = zeros (Nx, Ny) ;
for px=1:Nx
par f or py=1:Ny
del taT = Tn( px , py , : ) Tref ;
[ val i dx ] = max( del taT ( : ) ) ;
Cl assi cal Maxi gram ( px , py) = val ;
Cl assi cal Ti megram ( px , py) = i dx ;
end
end
%%
%% No Ref erence Approach
rfMaxigram = zeros (Nx, Ny) ;
rfTimegram = zeros (Nx, Ny) ;
Tn = permute (Tn, [ 3 1 2 ] ) ;
117
t = [ 1 : Nt ] ;
that = ( t mean( t ( : ) ) ) / std( t ( : ) ) ;
for px=1:Nx
par f or py=1:Ny
c o e f f 3 = pol yf i t ( t , Tn( : , px , py) , 3) ;
Tl = polyval ( c oe f f 3 , t ) ;
c o e f f 6 = pol yf i t ( that , Tn( : , px , py) , 6) ;
Th = polyval ( c oe f f 6 , that ) ;
del taT = Th Tl ;
[ val i dx ] = max( del taT ( : ) ) ;
rfMaxigram( px , py) = val ;
rfTimegram( px , py) = i dx ;
end
end
%% Output
outCube ( : , : , 1 ) = imnorm( Cl assi cal Maxi gram ) ;
outCube ( : , : , 2 ) = imnorm( Cl assi cal Ti megram ) ;
outTags {1} = Cl a s s i c a l Maxigram ;
outTags {2} = Cl a s s i c a l Timegram ;
outCube ( : , : , 3 ) = imnorm( rfMaxigram) ;
outCube ( : , : , 4 ) = imnorm( rfTimegram) ;
outTags {3} = Ref er ence Free Maxigram ;
outTags {4} = Ref er ence Free Timegram ;
118
Di spl ayTi me ( toc ) ;
f pri ntf ( Dynamic Thermal Tomography (DTT) Al gori thm Complete ! \ n )
;
B.5 Pulse Phase Thermography (PPT)
function [ outCube outTags ampCube phaseCube ] = ppt ( imageCube )
% PPT Pul se Phase Thermography al gor i t hm f or pul s ed thermography
% dat a . The al gor i t hm uses anal og i e s t o modul ated thermography
and t he
% Fouri er t ransf orm t o e v al uat e t he ampl i t ude and phase images of
t he
% dat a .
%
% [Y TAGS AMPCUBE PHASECUBE] = PPT(X) Process 3D image cube X
and r e s u l t s
% i n Y ( f i n i s h de s c r i pt i on ) . AMPCUBE and PHASECUBE are image
cubes
% r e pr e s e nt i ng t he ampl i t ude and phase responses , r e s p e c t i v e l y . X
, AMPCUBE,
% and PHASECUBE are U by V by W, where U and V are t he di mensi ons
of each
% image frame and W i s t he number of f rames . TAGS i s a vect or of
c e l l
% arrays cont ai ni ng a t e x t de s c r i pt i on of each out put frame .
%
% Ref erence Paper ( s ) :
% 1) X. Maldague and S. Mari net t i , Pul se phase i nf r ar e d
thermography ,
119
% Appl i ed Physi cs , 1996.
%
% 2) X. P. Maldague , Theory and Pr act i ce of i nf r ar e d t echnol ogy
f or nonde
% s t r u c t i v e t e s t i ng . John Wil ey & Sons , Inc . , 2001.
%
%
% See al s o POLYSMOOTH, FFT.
%
% PULSED THERMOGRAPHY PROCESSING TOOLBOX
% Author : Cory Andrew Larsen
f pri ntf ( Pr oces s i ng wi th Pul se Phase Thermography . . . \ n ) ;
ti c ;
% f i nd di mensi ons
[ Nx Ny Nt ] = si ze ( imageCube ) ;
M=NxNy;
imageMat = reshape( imageCube ,M, Nt ) ;
% Take Fouri er t ransf orm
F = zeros (M, Nt ) ;
par f or nn=1:M
F( nn , : ) = f f t ( imageMat ( nn , : ) ) ;
end
% Find Ampl i tude & Phase images
ampCube = reshape( abs (F) , Nx, Ny, Nt ) ;
120
phaseCube = reshape( angle (F) , Nx, Ny, Nt ) ;
outCube = imnorm( phaseCube ) ;
outTags = genNumTags ( outCube ) ;
Di spl ayTi me ( toc ) ;
f pri ntf ( Pul se Phase Thermography (PPT) Complete ! \ n ) ;
B.6 Time-Dierence Contrast (TDC)
function [ outCube outTags ] = Ti meDi f f er ence ( imageCube )
% TIMEDIFFERENCE Subt r act s a ti me d i f f e r e nc e bet ween every f i f t h
image i n a
% image cube by f i r s t perf ormi ng a l e as t s quares f i t f or t he
image
% i n t e n s i t i e s .
%
% [Y TAGS] = TIMEDIFFERENCE(X) Process 3D image cube X and
r e s u l t s i n t he
% proces s ed out put cube Y. TAGS i s a vect or of c e l l arrays
cont ai ni ng a
% t e x t de s c r i pt i on of each out put frame .
%
% Ref erence ( s ) :
% Larsen , Cory A. Document Fl ash Thermography . Master s Thesi s .
Utah St at e
% Uni ver s i t y . 2011.
%
%
121
% See al s o POLYSMOOTH, FFT.
%
% PULSED THERMOGRAPHY PROCESSING TOOLBOX
% Author : Cory Andrew Larsen
f pri ntf ( Al gori thm Time Di f f e r e nc e Begi nni ng . . . \ n ) ;
ti c ;
addpath ( . . / l i b / ) ;
[ Nx Ny Nt ] = si ze ( imageCube ) ;
opt i ons = opti mset ( Di spl ay , o f f ) ;
I 1 = imageCube ( : , : , 1 ) ;
outCube = zeros (Nx, Ny, Nt ) ;
par f or nn=1:Nt
I 2 = imageCube ( : , : , nn) ;
% Perform Mi ni mi zat i on
hfun = @( al pha ) TDfun( al pha , I1 , I 2 ) ;
al pha = fmi nunc ( hfun , 1 , opt i ons ) ;
% Remove Negat i ve Val ues and Cont rast St r e t c h
I d i f f = al pha I 2 I 1 ;
method = 2;
i f ( method == 1)
I out = I d i f f ;
I out ( I d i f f <= 0) = 0;
outCube ( : , : , nn) = imnorm( I out ) ;
el se
122
outCube ( : , : , nn) = I 2+I d i f f al pha ;
end
end
outTags = genNumTags ( outCube ) ;
outCube = f i xCube ( outCube ) ;
toc
Di spl ayTi me ( toc ) ;
f pri ntf ( Time Di f f e r e nc e Complete ! \ n ) ;
end
function [ f ] = TDfun( al pha , I1 , I 2 )
f = norm( al pha I2I 1 ) 2;
end
B.7 Total Harmonic Distortion (THD)
function [ outCube outTags ] = THD( imageCube )
% THD Tot al Harmonic Di s t or t i on al gor i t hm f or pul s ed thermography
% dat a . The al gor i t hm enhances de f e c t c ont r as t by t ransf ormi ng
each p i x e l
% sequence i n t he Fouri er domain and e v al uat i ng t he amount of
noi s e
% wi t hi n t he s i g na l which i s i nd i c a t i v e of a de f e c t .
%
% [Y TAGS] = THD(X) Process 3D image cube X and r e s u l t s
123
% i n imageCube Y. Y cont ai ns s e v e r al images i nc l udi ng THD power
d e f i ni t i o n
% (THDp) , THD v ol t ag e d e f i ni t i o n (THDa) , and a s i gnal tonoi s e
r at i o (SNR)
% image . TAGS i s a c e l l array cont ai ni ng s t r i ng d e s c r i p t i o ns of
t he out put
% images cont ai ned i n Y.
%
%
% Ref erence ( s ) :
% Larsen , Cory A. Document Fl ash Thermography . Master s Thesi s .
Utah St at e
% Uni ver s i t y . 2011.
%
% See al s o FFT.
%
% PULSED THERMOGRAPHY PROCESSING TOOLBOX
% Author : Cory Andrew Larsen
f pri ntf ( Al gori thm THD Begi nni ng . . . \ n ) ;
ti c ;
[ Nx Ny Nt ] = si ze ( imageCube ) ;
M=NxNy;
imageMat = reshape( imageCube , [M Nt ] ) ;
% Find Center of Si gnal
i f (mod( Nt , 2 ) )
124
C = cei l ( Nt /2) ;
el se
C = cei l ( Nt /2) +1;
end
Fpeak = zeros ( 1 ,M) ;
THDp = zeros ( 1 ,M) ;
THDa = zeros ( 1 ,M) ;
SNR = zeros ( 1 ,M) ;
par f or nn=1:M
% Find Fouri er Transform
I = imageMat ( nn , : ) ;
F = f f t s hi f t ( abs ( f f t ( I ) ) ) ;
% Peak Val ue
Fpeak ( nn) = max(F( : ) ) ;
% Cal cul at e Power THD
P1 = F(C) 2;
Ptot = sum(F(C+1:end) . 2) ;
THDp( nn) = Ptot /P1 ;
% Cal cul at e Ampl i tude THD
THDa( nn) = sqrt ( Ptot ) / F(C) ;
% Cal cul at e SNR???
SNR( nn) = mean(F( : ) ) /std(F( : ) ) ;
125
end
outCube ( : , : , 4 ) = reshape( Fpeak , Nx, Ny) ;
outTags {4} = Peak FFT Magnitude ;
outCube ( : , : , 1 ) = reshape(THDp, Nx, Ny) ;
outTags {1} = THDp ;
outCube ( : , : , 2 ) = reshape(THDa, Nx, Ny) ;
outTags {2} = THDa ;
outCube ( : , : , 3 ) = reshape(SNR, Nx, Ny) ;
outTags {3} = SNR ;
outCube = imnorm( outCube ) ;
f pri ntf ( Al gori thm THD Fi ni s hed ! \ n ) ;
toc
Di spl ayTi me ( toc ) ;
B.8 Markov Error Contrast (MEC)
function [ outCube outTags ] = MarkovError ( imageCube )
% MEC Markov Error Cont rast al gor i t hm f or pul s ed thermography
% dat a . The al gor i t hm enhances de f e c t c ont r as t by model i ng each
pi x e l s
% t emporal sequence as a Markov chai n .
%
% [Y TAGS] = MEC(X) Process 3D image cube X and r e s u l t s
126
% i n imageCube Y. Y cont ai ns t he r e s u l t i ng c ont r as t images
cr eat ed from
% d i f f e r e nc i ng t he sequence from t he Markov chai n es t i mat i on .
TAGS i s a
% c e l l array cont ai ni ng s t r i ng de s c r i p t i o ns of t he out put images
cont ai ned
% i n Y.
%
%
% Ref erence ( s ) :
% Larsen , Cory A. Document Fl ash Thermography . Master s Thesi s .
Utah St at e
% Uni ver s i t y . 2011.
%
%
% PULSED THERMOGRAPHY PROCESSING TOOLBOX
% Author : Cory Andrew Larsen
ti c ;
[ Nx Ny Nt ] = si ze ( imageCube ) ;
M=NxNy;
imageMat = reshape( imageCube , [M Nt ] ) ;
outCube = zeros (M, Nt ) ;
par f or nn=1:M
T = zeros ( 1 , Nt ) ;
I = imageMat ( nn , : ) ;
% Cal cul at e Markov Est i mat e Chain
T( 1) = I ( 1) ;
127
for t t =2:Nt
T( t t ) = I ( tt 1)sqrt ( ( tt 1)/ t t ) ;
end
outCube ( nn , : ) = I T; ;
end
outCube = imnorm( reshape( outCube , [ Nx Ny Nt ] ) ) ;
outTags = genNumTags ( outCube ) ;
Di spl ayTi me ( toc ) ;
B.9 Time Constant Analysis (TCA)
function [ outCube outTag ] = t ca ( imageCube )
% TCA Fi nds t he ti me cons t ant of t he ex pone nt i al decay of t he
t hermal decay
% s i g na l .
%
% [Y TAGS] = TCA(X) Process 3D image cube X and r e s u l t s i n Y,
c ons i s t i ng
% of two frames , one cont ai ni ng t he ampl i t ude c o e f f i c i e n t and t he
ot her
% cont ai ni ng t he ti me cons t ant image . X i s U by V by W, where U
and V are
% t he di mensi ons of each image frame and W i s t he number of
f rames . Y i s
% U by V by 2. TAGS i s a vect or of c e l l arrays cont ai ni ng a t e x t
% de s c r i pt i on of each out put frame .
128
%
% Ref erence ( s ) :
% Larsen , Cory A. Document Fl ash Thermography . Master s Thesi s .
Utah St at e
% Uni ver s i t y . 2011.
%
% See al s o SVD, FIXCUBE.
%
% PULSED THERMOGRAPHY PROCESSING TOOLBOX
% Author : Cory Andrew Larsen
ti c ;
% f i nd di mensi ons
[ Nx Ny Nt ] = si ze ( imageCube ) ;
M = NxNy;
imageMat = reshape( imageCube , [M Nt ] ) ;
% I n i t i a l i z e Fi t model
model= aexp(x/b)+c ;
FT = f i t t y pe ( model ) ;
t = [ 1 : Nt ] ;
% I n i t i a l i z e Mat ri ces
Tau = zeros ( 1 ,M) ;
A = zeros ( 1 ,M) ;
% C = z er os (1 ,M) ;
par f or nn=1:M
I = imageMat ( nn , : ) ;
129
% Perform Least Squares Fi t
opt i ons = f i t o p t i o n s ( Method , Nonl i near Leas t Squar es ) ;
opt i ons . Lower = [ 0 1 0 ] ;
opt i ons . St ar t Poi nt = [ . 5 7 0 ] ;
F = f i t ( t , I , FT, opt i ons ) ;
% Get System Parameters
cv = c o e f f v a l ue s (F) ;
% Save val ues
Tau( nn) = cv ( 2) ;
A( nn) = cv ( 1) ;
% C(nn) = cv (3) ;
end
outCube ( : , : , 1 ) = reshape(A, Nx, Ny) ;
outTag{1} = Amplitude ;
outCube ( : , : , 2 ) = reshape(Tau , Nx, Ny) ;
outTag{2} = Time Constant ;
outCube = imnorm( outCube ) ;
Di spl ayTi me ( toc ) ;
B.10 Matched Filters
function [ outCube outTags ] = Mat chedFi l t er s ( imageCube )
% MATCHEDFILTERS Appl i es f our matched f i l t e r d e f i nt i o ns t o f l a s h
% thermography dat a s e t .
%
130
% [Y TAGS] = MATCHEDFILTERS(X) Process 3D image cube X and
r e s u l t s
% i n imageCube Y. Y cont ai ns t he images cr eat ed from t he f our
matched
% f i l t e r d e f i n i t i o n s . TAGS i s a c e l l array cont ai ni ng s t r i ng
d e s c r i p t i o ns
% of t he out put images cont ai ned i n Y.
%
%
% Ref erence ( s ) :
% Larsen , Cory A. Document Fl ash Thermography . Master s Thesi s .
Utah St at e
% Uni ver s i t y . 2011.
%
%
% PULSED THERMOGRAPHY PROCESSING TOOLBOX
% Author : Cory Andrew Larsen
f pri ntf ( Running Matched Fi l t e r s . . . \ n ) ;
[ Nx Ny Nt ] = si ze ( imageCube ) ;
% Se l e c t Pi x e l s from Image and cr eat e H, B, and U mat ri ces %
NumPts = 5;
f pri ntf ( Choose Target Pi xe l s : ) ;
h = Se l e c t Poi nt s (NumPts , imageCube , 1 ) ;
pause ( 1)
f pri ntf ( Choose Background Pi xe l s : ) ;
b = Se l e c t Poi nt s (NumPts , imageCube , 1 ) ;
131
ti c
s = hb ;
s t = s ;
imageCube = permute ( imageCube , [ 3 1 2 ] ) ;
% Cal c . Data Vari ances
M = NxNy;
N = Nt ;
A = reshape( imageCube ,M, N) ;
Rinv = inv( cov(A) ) ;
%%
rho = 1/ sqrt ( s t Rinv s ) ;
mean s = mean( s ( : ) ) ;
var s = var ( s ( : ) ) ;
const 1 = sqrt ( s t Rinv s ) ;
const 3 = sqrt ( s t s ) ;
% I n i t i a l i z e Arrays
SAM = zeros (Nx, Ny) ;
ACE = zeros (Nx, Ny) ;
t s t a t = zeros (Nx, Ny) ; Fstat = zeros (Nx, Ny) ;
%% Cal cul at e Matched Fi l t e r Res ul t s
for px=1:Nx
for py=1:Ny
132
x = imageCube ( : , px , py) ;
xt = x ;
const 2 = xt Rinvx ;
% Spe c t r al Angl e Map (SAM)
SMF = s t x ;
SAM( px , py) = SMF / ( const 3 sqrt ( xt x) ) ;
% Cl ut t e r Matched Fi l t e r (CMF)
CMF = s t Rinvx ;
% Adapt i ve Coherence Est i mat or (ACE)
ACE( px , py) = CMF / ( const 1 sqrt ( const 2 ) ) ;
% t s t a t i s t i c
d = mean s mean( x ( : ) ) / ( sqrt ( var s + var ( x ( : ) ) ) /2 )
;
t s t a t ( px , py) = CMF rhosqrt ( d1) / sqrt ( const 2 rho
2CMF2 ) ;
% F s t a t i s t i c
Fstat ( px , py) = t s t a t ( px , py) 2;
%%
end
end
% St ore Output %
outCube ( : , : , 1 ) = imnorm(SAM) ;
outTags {1} = MFSAM ;
133
outCube ( : , : , 2 ) = imnorm(ACE) ;
outTags {2} = MFACE ;
outCube ( : , : , 3 ) = imnorm( abs ( t s t a t ) ) ;
outTags {3} = MFt S t a t i s t i c ;
outCube ( : , : , 4 ) = imnorm( Fstat ) ;
outTags {5} = MfF s t a t i s t i c ;
%%
outCube = imnorm( outCube ) ;
f pri ntf ( Matched Fi l t e r s Pr oces s i ng Complete ! \ n ) ;
Di spl ayTi me ( toc ) ;
end